{
"pdf_info": [
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
108,
79,
507,
116
],
"lines": [
{
"bbox": [
105,
78,
508,
98
],
"spans": [
{
"bbox": [
105,
78,
508,
98
],
"score": 1.0,
"content": "PRACTICAL ADVERSARIAL ATTACKS ON BRAIN–",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
98,
289,
117
],
"spans": [
{
"bbox": [
106,
98,
289,
117
],
"score": 1.0,
"content": "COMPUTER INTERFACES",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
112,
135,
244,
157
],
"lines": [
{
"bbox": [
113,
135,
201,
147
],
"spans": [
{
"bbox": [
113,
135,
201,
147
],
"score": 1.0,
"content": "Anonymous authors",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
111,
146,
245,
158
],
"spans": [
{
"bbox": [
111,
146,
245,
158
],
"score": 1.0,
"content": "Paper under double-blind review",
"type": "text"
}
],
"index": 3
}
],
"index": 2.5
},
{
"type": "title",
"bbox": [
278,
186,
333,
199
],
"lines": [
{
"bbox": [
277,
186,
335,
200
],
"spans": [
{
"bbox": [
277,
186,
335,
200
],
"score": 1.0,
"content": "ABSTRACT",
"type": "text"
}
],
"index": 4
}
],
"index": 4
},
{
"type": "text",
"bbox": [
143,
209,
469,
418
],
"lines": [
{
"bbox": [
141,
210,
469,
222
],
"spans": [
{
"bbox": [
141,
210,
469,
222
],
"score": 1.0,
"content": "Deep learning has been widely employed in brain–computer interfaces (BCIs) to",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
141,
221,
470,
234
],
"spans": [
{
"bbox": [
141,
221,
470,
234
],
"score": 1.0,
"content": "decode a subject’s intentions based on recorded brain activities enabling direct",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
141,
232,
469,
244
],
"spans": [
{
"bbox": [
141,
232,
469,
244
],
"score": 1.0,
"content": "interaction with computers and machines. BCI systems play a crucial role in",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
142,
244,
470,
255
],
"spans": [
{
"bbox": [
142,
244,
470,
255
],
"score": 1.0,
"content": "motor rehabilitation and have recently experienced a significant market boost as",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
141,
254,
470,
266
],
"spans": [
{
"bbox": [
141,
254,
470,
266
],
"score": 1.0,
"content": "consumer-grade products. Recent studies have shown that deep learning-based",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
141,
265,
469,
277
],
"spans": [
{
"bbox": [
141,
265,
469,
277
],
"score": 1.0,
"content": "BCIs are vulnerable to adversarial attacks. Failures in such systems might cause",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
142,
276,
470,
288
],
"spans": [
{
"bbox": [
142,
276,
470,
288
],
"score": 1.0,
"content": "medical misdiagnoses, physical harm, and financial damages, hence it is of utmost",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
142,
288,
469,
299
],
"spans": [
{
"bbox": [
142,
288,
469,
299
],
"score": 1.0,
"content": "importance to analyze and understand in-depth, potential malicious attacks to",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
142,
298,
469,
310
],
"spans": [
{
"bbox": [
142,
298,
469,
310
],
"score": 1.0,
"content": "develop countermeasures. In this work, we present the first study that analyzes",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
142,
309,
470,
321
],
"spans": [
{
"bbox": [
142,
309,
470,
321
],
"score": 1.0,
"content": "and models adversarial attacks based on physical domain constraints in EEG-",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
141,
320,
469,
332
],
"spans": [
{
"bbox": [
141,
320,
469,
332
],
"score": 1.0,
"content": "based BCIs. Specifically, we assess the robustness of EEGNet which is the",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
141,
331,
469,
343
],
"spans": [
{
"bbox": [
141,
331,
469,
343
],
"score": 1.0,
"content": "current state-of-the-art network for embedded BCIs. We propose new methods",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
141,
342,
469,
354
],
"spans": [
{
"bbox": [
141,
342,
469,
354
],
"score": 1.0,
"content": "to induce denial-of-service attacks and incorporate domain-specific insights and",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
141,
353,
470,
365
],
"spans": [
{
"bbox": [
141,
353,
470,
365
],
"score": 1.0,
"content": "constraints to accomplish two key goals: (i) create smooth adversarial attacks",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
141,
364,
470,
375
],
"spans": [
{
"bbox": [
141,
364,
470,
375
],
"score": 1.0,
"content": "that are physiologically plausible; (ii) consider the realistic case where the attack",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
141,
375,
470,
387
],
"spans": [
{
"bbox": [
141,
375,
470,
387
],
"score": 1.0,
"content": "happens at the origin of the signal acquisition and it propagates on the human head.",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
141,
385,
470,
398
],
"spans": [
{
"bbox": [
141,
385,
470,
398
],
"score": 1.0,
"content": "Our results show that EEGNet is significantly vulnerable to adversarial attacks",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
141,
396,
469,
408
],
"spans": [
{
"bbox": [
141,
396,
308,
408
],
"score": 1.0,
"content": "with an attack success rate of more than",
"type": "text"
},
{
"bbox": [
309,
397,
329,
407
],
"score": 0.86,
"content": "50 \\%",
"type": "inline_equation"
},
{
"bbox": [
329,
396,
469,
408
],
"score": 1.0,
"content": ". With our work, we want to raise",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
141,
407,
443,
420
],
"spans": [
{
"bbox": [
141,
407,
443,
420
],
"score": 1.0,
"content": "awareness and incentivize future developments of proper countermeasures.",
"type": "text"
}
],
"index": 23
}
],
"index": 14
},
{
"type": "title",
"bbox": [
108,
438,
206,
450
],
"lines": [
{
"bbox": [
105,
437,
208,
453
],
"spans": [
{
"bbox": [
105,
437,
208,
453
],
"score": 1.0,
"content": "1 INTRODUCTION",
"type": "text"
}
],
"index": 24
}
],
"index": 24
},
{
"type": "text",
"bbox": [
107,
463,
505,
616
],
"lines": [
{
"bbox": [
105,
461,
505,
476
],
"spans": [
{
"bbox": [
105,
461,
505,
476
],
"score": 1.0,
"content": "Recent work has shown that adversarial perturbations can cause state-of-the-art (SoA) deep learning",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
474,
506,
485
],
"spans": [
{
"bbox": [
106,
474,
506,
485
],
"score": 1.0,
"content": "models to misbehave in various domains including vision (Szegedy et al., 2014; Goodfellow et al.,",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
484,
506,
497
],
"spans": [
{
"bbox": [
106,
484,
506,
497
],
"score": 1.0,
"content": "2015), NLP (Li et al., 2019a; Zhang et al., 2020), speech (Qin et al., 2019; Li et al., 2019b), and",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
495,
505,
508
],
"spans": [
{
"bbox": [
106,
495,
505,
508
],
"score": 1.0,
"content": "biomedicine (Finlayson et al., 2019; Han et al., 2020). Neural networks have been applied in",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
506,
506,
519
],
"spans": [
{
"bbox": [
105,
506,
506,
519
],
"score": 1.0,
"content": "brain–computer interfaces (BCIs) achieving impressive results (Lawhern et al., 2018; Dose et al.,",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
516,
506,
530
],
"spans": [
{
"bbox": [
105,
516,
506,
530
],
"score": 1.0,
"content": "2018). A BCI enables direct interactions with external devices based on brain activities, typically",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
528,
505,
541
],
"spans": [
{
"bbox": [
105,
528,
505,
541
],
"score": 1.0,
"content": "recorded using electroencephalographic (EEG) systems. It can provide a communication pathway",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
539,
505,
552
],
"spans": [
{
"bbox": [
106,
539,
505,
552
],
"score": 1.0,
"content": "for severely paralyzed patients or assist in rehabilitation (Chaudhary et al., 2016). Besides medical",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
550,
505,
563
],
"spans": [
{
"bbox": [
105,
550,
505,
563
],
"score": 1.0,
"content": "applications, recent developments in wearable devices have pushed BCIs towards consumer-grade",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
561,
506,
574
],
"spans": [
{
"bbox": [
106,
561,
506,
574
],
"score": 1.0,
"content": "products to improve life quality (Aricò et al., 2020), e.g., the Interaxon Muse headband for stress",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
573,
505,
584
],
"spans": [
{
"bbox": [
106,
573,
505,
584
],
"score": 1.0,
"content": "relief (Arsalan et al., 2019) or the Emotiv headset for controlling drones (Marin et al., 2020) and",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
104,
582,
506,
596
],
"spans": [
{
"bbox": [
104,
582,
506,
596
],
"score": 1.0,
"content": "ground vehicles (Zhuang et al., 2021). Safety in BCI systems is paramount (Dutta, 2020; Bernal et al.,",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
593,
507,
608
],
"spans": [
{
"bbox": [
105,
593,
507,
608
],
"score": 1.0,
"content": "2021), because a failure would cause misdiagnoses, user frustration, or even danger while driving a",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
604,
406,
618
],
"spans": [
{
"bbox": [
105,
604,
406,
618
],
"score": 1.0,
"content": "wheelchair or controlling a drone, causing physical and financial damages.",
"type": "text"
}
],
"index": 38
}
],
"index": 31.5
},
{
"type": "text",
"bbox": [
107,
622,
505,
732
],
"lines": [
{
"bbox": [
106,
621,
505,
634
],
"spans": [
{
"bbox": [
106,
621,
505,
634
],
"score": 1.0,
"content": "Zhang & Wu (2019) were the first to show that EEG-based BCIs are vulnerable to adversarial attacks",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
633,
505,
645
],
"spans": [
{
"bbox": [
105,
633,
505,
645
],
"score": 1.0,
"content": "by proposing an unsupervised fast gradient sign method (FGSM) (Goodfellow et al., 2015). More",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
644,
506,
657
],
"spans": [
{
"bbox": [
105,
644,
506,
657
],
"score": 1.0,
"content": "recent work has proposed a more practical attack where a universal adversarial perturbation (UAP)",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
654,
505,
668
],
"spans": [
{
"bbox": [
105,
654,
505,
668
],
"score": 1.0,
"content": "is computed once and can be applied to all EEG trials without learning it for every new input (Liu",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
104,
664,
507,
680
],
"spans": [
{
"bbox": [
104,
664,
507,
680
],
"score": 1.0,
"content": "et al., 2021). Both works assume that the acquired signals are sent to a remote compute engine, e.g.,",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
104,
676,
506,
691
],
"spans": [
{
"bbox": [
104,
676,
506,
691
],
"score": 1.0,
"content": "a computer, and the attacker can alter the signals during the transmission by attaching a “jamming”",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
687,
506,
702
],
"spans": [
{
"bbox": [
105,
687,
506,
702
],
"score": 1.0,
"content": "module between the signal preprocessing step and the classifier. Recent developments in smart edge",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
698,
507,
712
],
"spans": [
{
"bbox": [
105,
698,
507,
712
],
"score": 1.0,
"content": "computing (Akmandor & Jha, 2018; Beach et al., 2021) eliminate the need for data transmission,",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
106,
710,
506,
721
],
"spans": [
{
"bbox": [
106,
710,
506,
721
],
"score": 1.0,
"content": "making this attack scenario inapplicable. Novel BCI solutions (Kartsch et al., 2019; Wang et al.,",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
720,
506,
733
],
"spans": [
{
"bbox": [
105,
720,
506,
733
],
"score": 1.0,
"content": "2020) embed the signal processing and classification directly at the sensor edge. A more practical",
"type": "text"
}
],
"index": 48
}
],
"index": 43.5
}
],
"page_idx": 0,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
308,
760
],
"lines": [
{
"bbox": [
302,
751,
308,
761
],
"spans": [
{
"bbox": [
302,
751,
308,
761
],
"score": 1.0,
"content": "1",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
108,
79,
507,
116
],
"lines": [
{
"bbox": [
105,
78,
508,
98
],
"spans": [
{
"bbox": [
105,
78,
508,
98
],
"score": 1.0,
"content": "PRACTICAL ADVERSARIAL ATTACKS ON BRAIN–",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
98,
289,
117
],
"spans": [
{
"bbox": [
106,
98,
289,
117
],
"score": 1.0,
"content": "COMPUTER INTERFACES",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
112,
135,
244,
157
],
"lines": [
{
"bbox": [
113,
135,
201,
147
],
"spans": [
{
"bbox": [
113,
135,
201,
147
],
"score": 1.0,
"content": "Anonymous authors",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
111,
146,
245,
158
],
"spans": [
{
"bbox": [
111,
146,
245,
158
],
"score": 1.0,
"content": "Paper under double-blind review",
"type": "text"
}
],
"index": 3
}
],
"index": 2.5,
"bbox_fs": [
111,
135,
245,
158
]
},
{
"type": "title",
"bbox": [
278,
186,
333,
199
],
"lines": [
{
"bbox": [
277,
186,
335,
200
],
"spans": [
{
"bbox": [
277,
186,
335,
200
],
"score": 1.0,
"content": "ABSTRACT",
"type": "text"
}
],
"index": 4
}
],
"index": 4
},
{
"type": "text",
"bbox": [
143,
209,
469,
418
],
"lines": [
{
"bbox": [
141,
210,
469,
222
],
"spans": [
{
"bbox": [
141,
210,
469,
222
],
"score": 1.0,
"content": "Deep learning has been widely employed in brain–computer interfaces (BCIs) to",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
141,
221,
470,
234
],
"spans": [
{
"bbox": [
141,
221,
470,
234
],
"score": 1.0,
"content": "decode a subject’s intentions based on recorded brain activities enabling direct",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
141,
232,
469,
244
],
"spans": [
{
"bbox": [
141,
232,
469,
244
],
"score": 1.0,
"content": "interaction with computers and machines. BCI systems play a crucial role in",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
142,
244,
470,
255
],
"spans": [
{
"bbox": [
142,
244,
470,
255
],
"score": 1.0,
"content": "motor rehabilitation and have recently experienced a significant market boost as",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
141,
254,
470,
266
],
"spans": [
{
"bbox": [
141,
254,
470,
266
],
"score": 1.0,
"content": "consumer-grade products. Recent studies have shown that deep learning-based",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
141,
265,
469,
277
],
"spans": [
{
"bbox": [
141,
265,
469,
277
],
"score": 1.0,
"content": "BCIs are vulnerable to adversarial attacks. Failures in such systems might cause",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
142,
276,
470,
288
],
"spans": [
{
"bbox": [
142,
276,
470,
288
],
"score": 1.0,
"content": "medical misdiagnoses, physical harm, and financial damages, hence it is of utmost",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
142,
288,
469,
299
],
"spans": [
{
"bbox": [
142,
288,
469,
299
],
"score": 1.0,
"content": "importance to analyze and understand in-depth, potential malicious attacks to",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
142,
298,
469,
310
],
"spans": [
{
"bbox": [
142,
298,
469,
310
],
"score": 1.0,
"content": "develop countermeasures. In this work, we present the first study that analyzes",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
142,
309,
470,
321
],
"spans": [
{
"bbox": [
142,
309,
470,
321
],
"score": 1.0,
"content": "and models adversarial attacks based on physical domain constraints in EEG-",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
141,
320,
469,
332
],
"spans": [
{
"bbox": [
141,
320,
469,
332
],
"score": 1.0,
"content": "based BCIs. Specifically, we assess the robustness of EEGNet which is the",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
141,
331,
469,
343
],
"spans": [
{
"bbox": [
141,
331,
469,
343
],
"score": 1.0,
"content": "current state-of-the-art network for embedded BCIs. We propose new methods",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
141,
342,
469,
354
],
"spans": [
{
"bbox": [
141,
342,
469,
354
],
"score": 1.0,
"content": "to induce denial-of-service attacks and incorporate domain-specific insights and",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
141,
353,
470,
365
],
"spans": [
{
"bbox": [
141,
353,
470,
365
],
"score": 1.0,
"content": "constraints to accomplish two key goals: (i) create smooth adversarial attacks",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
141,
364,
470,
375
],
"spans": [
{
"bbox": [
141,
364,
470,
375
],
"score": 1.0,
"content": "that are physiologically plausible; (ii) consider the realistic case where the attack",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
141,
375,
470,
387
],
"spans": [
{
"bbox": [
141,
375,
470,
387
],
"score": 1.0,
"content": "happens at the origin of the signal acquisition and it propagates on the human head.",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
141,
385,
470,
398
],
"spans": [
{
"bbox": [
141,
385,
470,
398
],
"score": 1.0,
"content": "Our results show that EEGNet is significantly vulnerable to adversarial attacks",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
141,
396,
469,
408
],
"spans": [
{
"bbox": [
141,
396,
308,
408
],
"score": 1.0,
"content": "with an attack success rate of more than",
"type": "text"
},
{
"bbox": [
309,
397,
329,
407
],
"score": 0.86,
"content": "50 \\%",
"type": "inline_equation"
},
{
"bbox": [
329,
396,
469,
408
],
"score": 1.0,
"content": ". With our work, we want to raise",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
141,
407,
443,
420
],
"spans": [
{
"bbox": [
141,
407,
443,
420
],
"score": 1.0,
"content": "awareness and incentivize future developments of proper countermeasures.",
"type": "text"
}
],
"index": 23
}
],
"index": 14,
"bbox_fs": [
141,
210,
470,
420
]
},
{
"type": "title",
"bbox": [
108,
438,
206,
450
],
"lines": [
{
"bbox": [
105,
437,
208,
453
],
"spans": [
{
"bbox": [
105,
437,
208,
453
],
"score": 1.0,
"content": "1 INTRODUCTION",
"type": "text"
}
],
"index": 24
}
],
"index": 24
},
{
"type": "text",
"bbox": [
107,
463,
505,
616
],
"lines": [
{
"bbox": [
105,
461,
505,
476
],
"spans": [
{
"bbox": [
105,
461,
505,
476
],
"score": 1.0,
"content": "Recent work has shown that adversarial perturbations can cause state-of-the-art (SoA) deep learning",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
474,
506,
485
],
"spans": [
{
"bbox": [
106,
474,
506,
485
],
"score": 1.0,
"content": "models to misbehave in various domains including vision (Szegedy et al., 2014; Goodfellow et al.,",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
484,
506,
497
],
"spans": [
{
"bbox": [
106,
484,
506,
497
],
"score": 1.0,
"content": "2015), NLP (Li et al., 2019a; Zhang et al., 2020), speech (Qin et al., 2019; Li et al., 2019b), and",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
495,
505,
508
],
"spans": [
{
"bbox": [
106,
495,
505,
508
],
"score": 1.0,
"content": "biomedicine (Finlayson et al., 2019; Han et al., 2020). Neural networks have been applied in",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
506,
506,
519
],
"spans": [
{
"bbox": [
105,
506,
506,
519
],
"score": 1.0,
"content": "brain–computer interfaces (BCIs) achieving impressive results (Lawhern et al., 2018; Dose et al.,",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
516,
506,
530
],
"spans": [
{
"bbox": [
105,
516,
506,
530
],
"score": 1.0,
"content": "2018). A BCI enables direct interactions with external devices based on brain activities, typically",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
528,
505,
541
],
"spans": [
{
"bbox": [
105,
528,
505,
541
],
"score": 1.0,
"content": "recorded using electroencephalographic (EEG) systems. It can provide a communication pathway",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
539,
505,
552
],
"spans": [
{
"bbox": [
106,
539,
505,
552
],
"score": 1.0,
"content": "for severely paralyzed patients or assist in rehabilitation (Chaudhary et al., 2016). Besides medical",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
550,
505,
563
],
"spans": [
{
"bbox": [
105,
550,
505,
563
],
"score": 1.0,
"content": "applications, recent developments in wearable devices have pushed BCIs towards consumer-grade",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
561,
506,
574
],
"spans": [
{
"bbox": [
106,
561,
506,
574
],
"score": 1.0,
"content": "products to improve life quality (Aricò et al., 2020), e.g., the Interaxon Muse headband for stress",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
573,
505,
584
],
"spans": [
{
"bbox": [
106,
573,
505,
584
],
"score": 1.0,
"content": "relief (Arsalan et al., 2019) or the Emotiv headset for controlling drones (Marin et al., 2020) and",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
104,
582,
506,
596
],
"spans": [
{
"bbox": [
104,
582,
506,
596
],
"score": 1.0,
"content": "ground vehicles (Zhuang et al., 2021). Safety in BCI systems is paramount (Dutta, 2020; Bernal et al.,",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
593,
507,
608
],
"spans": [
{
"bbox": [
105,
593,
507,
608
],
"score": 1.0,
"content": "2021), because a failure would cause misdiagnoses, user frustration, or even danger while driving a",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
604,
406,
618
],
"spans": [
{
"bbox": [
105,
604,
406,
618
],
"score": 1.0,
"content": "wheelchair or controlling a drone, causing physical and financial damages.",
"type": "text"
}
],
"index": 38
}
],
"index": 31.5,
"bbox_fs": [
104,
461,
507,
618
]
},
{
"type": "text",
"bbox": [
107,
622,
505,
732
],
"lines": [
{
"bbox": [
106,
621,
505,
634
],
"spans": [
{
"bbox": [
106,
621,
505,
634
],
"score": 1.0,
"content": "Zhang & Wu (2019) were the first to show that EEG-based BCIs are vulnerable to adversarial attacks",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
633,
505,
645
],
"spans": [
{
"bbox": [
105,
633,
505,
645
],
"score": 1.0,
"content": "by proposing an unsupervised fast gradient sign method (FGSM) (Goodfellow et al., 2015). More",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
644,
506,
657
],
"spans": [
{
"bbox": [
105,
644,
506,
657
],
"score": 1.0,
"content": "recent work has proposed a more practical attack where a universal adversarial perturbation (UAP)",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
654,
505,
668
],
"spans": [
{
"bbox": [
105,
654,
505,
668
],
"score": 1.0,
"content": "is computed once and can be applied to all EEG trials without learning it for every new input (Liu",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
104,
664,
507,
680
],
"spans": [
{
"bbox": [
104,
664,
507,
680
],
"score": 1.0,
"content": "et al., 2021). Both works assume that the acquired signals are sent to a remote compute engine, e.g.,",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
104,
676,
506,
691
],
"spans": [
{
"bbox": [
104,
676,
506,
691
],
"score": 1.0,
"content": "a computer, and the attacker can alter the signals during the transmission by attaching a “jamming”",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
687,
506,
702
],
"spans": [
{
"bbox": [
105,
687,
506,
702
],
"score": 1.0,
"content": "module between the signal preprocessing step and the classifier. Recent developments in smart edge",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
698,
507,
712
],
"spans": [
{
"bbox": [
105,
698,
507,
712
],
"score": 1.0,
"content": "computing (Akmandor & Jha, 2018; Beach et al., 2021) eliminate the need for data transmission,",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
106,
710,
506,
721
],
"spans": [
{
"bbox": [
106,
710,
506,
721
],
"score": 1.0,
"content": "making this attack scenario inapplicable. Novel BCI solutions (Kartsch et al., 2019; Wang et al.,",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
720,
506,
733
],
"spans": [
{
"bbox": [
105,
720,
506,
733
],
"score": 1.0,
"content": "2020) embed the signal processing and classification directly at the sensor edge. A more practical",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
106,
83,
505,
95
],
"spans": [
{
"bbox": [
106,
83,
505,
95
],
"score": 1.0,
"content": "adversarial example has been identified by Meng et al. (2021). It consists of a square-shaped signal",
"type": "text",
"cross_page": true
}
],
"index": 0
},
{
"bbox": [
105,
93,
505,
106
],
"spans": [
{
"bbox": [
105,
93,
505,
106
],
"score": 1.0,
"content": "that can be added to EEG trials before the preprocessing step. However, the attack is proposed as",
"type": "text",
"cross_page": true
}
],
"index": 1
},
{
"bbox": [
106,
105,
505,
116
],
"spans": [
{
"bbox": [
106,
105,
505,
116
],
"score": 1.0,
"content": "a backdoor key, which means that the attacker has direct access to the training dataset and pollutes",
"type": "text",
"cross_page": true
}
],
"index": 2
},
{
"bbox": [
105,
115,
505,
127
],
"spans": [
{
"bbox": [
105,
115,
505,
127
],
"score": 1.0,
"content": "it with adversarial examples, which is improbable if the attacker is not directly involved in the data",
"type": "text",
"cross_page": true
}
],
"index": 3
},
{
"bbox": [
105,
127,
505,
138
],
"spans": [
{
"bbox": [
105,
127,
505,
138
],
"score": 1.0,
"content": "acquisition or in the training of the classifier. Li et al. (2019b) have shown an attack scenario in the",
"type": "text",
"cross_page": true
}
],
"index": 4
},
{
"bbox": [
105,
138,
505,
150
],
"spans": [
{
"bbox": [
105,
138,
505,
150
],
"score": 1.0,
"content": "audio domain by considering the on-board edge processing of a wake-word detection system, where",
"type": "text",
"cross_page": true
}
],
"index": 5
},
{
"bbox": [
105,
149,
505,
160
],
"spans": [
{
"bbox": [
105,
149,
505,
160
],
"score": 1.0,
"content": "an adversarial audio trace is delivered to the environment causing denial-of-service (DoS). No similar",
"type": "text",
"cross_page": true
}
],
"index": 6
},
{
"bbox": [
105,
159,
308,
171
],
"spans": [
{
"bbox": [
105,
159,
308,
171
],
"score": 1.0,
"content": "studies can be currently found in the BCI domain.",
"type": "text",
"cross_page": true
}
],
"index": 7
}
],
"index": 43.5,
"bbox_fs": [
104,
621,
507,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
170
],
"lines": [
{
"bbox": [
106,
83,
505,
95
],
"spans": [
{
"bbox": [
106,
83,
505,
95
],
"score": 1.0,
"content": "adversarial example has been identified by Meng et al. (2021). It consists of a square-shaped signal",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
93,
505,
106
],
"spans": [
{
"bbox": [
105,
93,
505,
106
],
"score": 1.0,
"content": "that can be added to EEG trials before the preprocessing step. However, the attack is proposed as",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
105,
505,
116
],
"spans": [
{
"bbox": [
106,
105,
505,
116
],
"score": 1.0,
"content": "a backdoor key, which means that the attacker has direct access to the training dataset and pollutes",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
115,
505,
127
],
"spans": [
{
"bbox": [
105,
115,
505,
127
],
"score": 1.0,
"content": "it with adversarial examples, which is improbable if the attacker is not directly involved in the data",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
127,
505,
138
],
"spans": [
{
"bbox": [
105,
127,
505,
138
],
"score": 1.0,
"content": "acquisition or in the training of the classifier. Li et al. (2019b) have shown an attack scenario in the",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
138,
505,
150
],
"spans": [
{
"bbox": [
105,
138,
505,
150
],
"score": 1.0,
"content": "audio domain by considering the on-board edge processing of a wake-word detection system, where",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
149,
505,
160
],
"spans": [
{
"bbox": [
105,
149,
505,
160
],
"score": 1.0,
"content": "an adversarial audio trace is delivered to the environment causing denial-of-service (DoS). No similar",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
159,
308,
171
],
"spans": [
{
"bbox": [
105,
159,
308,
171
],
"score": 1.0,
"content": "studies can be currently found in the BCI domain.",
"type": "text"
}
],
"index": 7
}
],
"index": 3.5
},
{
"type": "text",
"bbox": [
107,
176,
505,
308
],
"lines": [
{
"bbox": [
106,
175,
506,
189
],
"spans": [
{
"bbox": [
106,
175,
506,
189
],
"score": 1.0,
"content": "Challenges: Designing natural attacks and modeling its propagation. Unlike in audio",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
187,
506,
200
],
"spans": [
{
"bbox": [
106,
187,
506,
200
],
"score": 1.0,
"content": "applications where the signal can simply propagate over-the-air and is sensed by a microphone, extra",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
198,
505,
210
],
"spans": [
{
"bbox": [
106,
198,
505,
210
],
"score": 1.0,
"content": "modeling is required to evaluate the signal propagation in BCIs based on the physical properties of the",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
208,
506,
223
],
"spans": [
{
"bbox": [
105,
208,
506,
223
],
"score": 1.0,
"content": "biological tissues. In this work, rather than assuming a “jamming” module between the preprocessing",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
220,
505,
232
],
"spans": [
{
"bbox": [
106,
220,
505,
232
],
"score": 1.0,
"content": "and the classification steps as in related works, we consider a more realistic and practically applicable",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
106,
231,
506,
243
],
"spans": [
{
"bbox": [
106,
231,
506,
243
],
"score": 1.0,
"content": "attack scenario where the adversarial perturbations are introduced at the source of the data acquisition,",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
243,
505,
254
],
"spans": [
{
"bbox": [
106,
243,
505,
254
],
"score": 1.0,
"content": "as showcased in Figure 6 in Appendix A. This can be achieved, for example, via electromagnetic",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
253,
505,
264
],
"spans": [
{
"bbox": [
105,
253,
505,
264
],
"score": 1.0,
"content": "waves delivered to the environment (Dutta, 2020) or via transcranial current stimulation with electrical",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
263,
506,
277
],
"spans": [
{
"bbox": [
105,
263,
506,
277
],
"score": 1.0,
"content": "current delivered directly to the scalp (Bodranghien et al., 2017; Fertonani et al., 2015), by exploiting",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
274,
506,
287
],
"spans": [
{
"bbox": [
105,
274,
506,
287
],
"score": 1.0,
"content": "wearable devices, such as smart glasses or over-ear headsets (Flowneuroscience, 2021; Marin et al.,",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
286,
506,
299
],
"spans": [
{
"bbox": [
106,
286,
506,
299
],
"score": 1.0,
"content": "2020). The adversarial perturbations translate into electrical signals propagating over the scalp and",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
297,
345,
309
],
"spans": [
{
"bbox": [
106,
297,
345,
309
],
"score": 1.0,
"content": "are sensed by the electrodes in addition to the EEG signals.",
"type": "text"
}
],
"index": 19
}
],
"index": 13.5
},
{
"type": "text",
"bbox": [
107,
313,
505,
390
],
"lines": [
{
"bbox": [
106,
313,
505,
326
],
"spans": [
{
"bbox": [
106,
313,
505,
326
],
"score": 1.0,
"content": "To guarantee the imperceptibility of the attacks, previous works in BCIs create perturbations that",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
325,
505,
336
],
"spans": [
{
"bbox": [
106,
325,
505,
336
],
"score": 1.0,
"content": "are small in amplitude (Zhang & Wu, 2019; Jiang et al., 2019; Liu et al., 2021), limiting the attack",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
106,
336,
506,
348
],
"spans": [
{
"bbox": [
106,
336,
506,
348
],
"score": 1.0,
"content": "success rate (ASR). Increased perturbation’s amplitude yields higher ASR (Meng et al., 2021), but",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
347,
506,
359
],
"spans": [
{
"bbox": [
106,
347,
506,
359
],
"score": 1.0,
"content": "makes the attack more easily detectable. Moreover, the generated perturbations are square-shaped,",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
358,
505,
370
],
"spans": [
{
"bbox": [
106,
358,
505,
370
],
"score": 1.0,
"content": "which is implausible for biosignals. Han et al. (2020) are the first to observe square-wave artifacts",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
367,
506,
382
],
"spans": [
{
"bbox": [
105,
367,
506,
382
],
"score": 1.0,
"content": "in biosignals’ attacks and propose smooth perturbations for electrocardiograms (ECGs). No similar",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
380,
243,
391
],
"spans": [
{
"bbox": [
106,
380,
243,
391
],
"score": 1.0,
"content": "works have been found for EEGs.",
"type": "text"
}
],
"index": 26
}
],
"index": 23
},
{
"type": "text",
"bbox": [
106,
396,
505,
517
],
"lines": [
{
"bbox": [
105,
395,
505,
409
],
"spans": [
{
"bbox": [
105,
395,
505,
409
],
"score": 1.0,
"content": "This work: Practical attacks on BCI models. To address the above technical challenges, and",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
407,
505,
420
],
"spans": [
{
"bbox": [
106,
407,
505,
420
],
"score": 1.0,
"content": "for analyzing the vulnerability of embedded BCI models in practical scenarios, we design a new",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
419,
505,
431
],
"spans": [
{
"bbox": [
106,
419,
505,
431
],
"score": 1.0,
"content": "attack algorithm that generates smooth adversarial examples based on the signals’ first derivative and",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
430,
504,
441
],
"spans": [
{
"bbox": [
106,
430,
504,
441
],
"score": 1.0,
"content": "model its propagation over the scalp based on a realistic head model by taking into consideration the",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
441,
505,
453
],
"spans": [
{
"bbox": [
106,
441,
505,
453
],
"score": 1.0,
"content": "attack source and the electrical and physical properties of the conducting tissues. This enables the",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
451,
506,
464
],
"spans": [
{
"bbox": [
105,
451,
506,
464
],
"score": 1.0,
"content": "creation of practically effective perturbations, that can be delivered by an external device to attack",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
461,
506,
474
],
"spans": [
{
"bbox": [
105,
461,
506,
474
],
"score": 1.0,
"content": "EEG-based BCIs at the source of signal acquisition. We attack the most energy-efficient network that",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
472,
506,
486
],
"spans": [
{
"bbox": [
105,
472,
506,
486
],
"score": 1.0,
"content": "has been embedded on microcontrollers for smart wearable BCIs called EEGNet (Lawhern et al.,",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
484,
504,
496
],
"spans": [
{
"bbox": [
106,
484,
504,
496
],
"score": 1.0,
"content": "2018; Schneider et al., 2020). It is a resource-friendly convolutional neural network (CNN) and is the",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
495,
506,
509
],
"spans": [
{
"bbox": [
105,
495,
506,
509
],
"score": 1.0,
"content": "SoA in terms of accuracy and energy-efficiency trade-off (Belwafi et al., 2018; Malekmohammadi",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
507,
325,
518
],
"spans": [
{
"bbox": [
106,
507,
325,
518
],
"score": 1.0,
"content": "et al., 2019; Wang et al., 2020; Schneider et al., 2020).",
"type": "text"
}
],
"index": 37
}
],
"index": 32
},
{
"type": "text",
"bbox": [
106,
523,
506,
676
],
"lines": [
{
"bbox": [
105,
522,
506,
535
],
"spans": [
{
"bbox": [
105,
522,
506,
535
],
"score": 1.0,
"content": "We evaluate our methods and show experimental results on BCIs based on the motor imagery (MI)",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
534,
506,
547
],
"spans": [
{
"bbox": [
105,
534,
506,
547
],
"score": 1.0,
"content": "paradigm, which is of special interest among others because it can be asynchronously self-paced",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
104,
543,
507,
559
],
"spans": [
{
"bbox": [
104,
543,
507,
559
],
"score": 1.0,
"content": "without external stimuli (Freer & Yang, 2020). By imagining the movement of different body parts,",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
555,
506,
569
],
"spans": [
{
"bbox": [
105,
555,
506,
569
],
"score": 1.0,
"content": "the decoded intention is translated into control signals. It is widely applied in several BCI applications,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
567,
506,
579
],
"spans": [
{
"bbox": [
105,
567,
506,
579
],
"score": 1.0,
"content": "such as the control of wheelchairss (Yu et al., 2018), prosthetic armss (Elstob & Secco, 2016), ground",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
578,
506,
590
],
"spans": [
{
"bbox": [
105,
578,
506,
590
],
"score": 1.0,
"content": "vehicles (Zhuang et al., 2021), and in communication (Brumberg et al., 2016). It has been proven",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
588,
506,
601
],
"spans": [
{
"bbox": [
105,
588,
506,
601
],
"score": 1.0,
"content": "to be the most difficult task to be attacked among the most common BCI paradigms (Zhang & Wu,",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
599,
506,
612
],
"spans": [
{
"bbox": [
105,
599,
506,
612
],
"score": 1.0,
"content": "2019; Meng et al., 2021). We evaluate our methods by “fooling” the victim model to always predict",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
610,
506,
623
],
"spans": [
{
"bbox": [
105,
610,
506,
623
],
"score": 1.0,
"content": "“rest” class. This essentially yields a DoS attack, because resting-state EEG signals are generally",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
621,
506,
633
],
"spans": [
{
"bbox": [
105,
621,
506,
633
],
"score": 1.0,
"content": "interpreted as no subject’s intention decoded, i.e., no control action needs to be taken by the BCI",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
632,
506,
645
],
"spans": [
{
"bbox": [
105,
632,
506,
645
],
"score": 1.0,
"content": "system (Yu et al., 2018). While for healthy subjects it might solely cause user frustration and financial",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
106,
643,
505,
655
],
"spans": [
{
"bbox": [
106,
643,
505,
655
],
"score": 1.0,
"content": "losses, for severely paralyzed patients it can lead to loss of communication and independence. We",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
654,
506,
667
],
"spans": [
{
"bbox": [
105,
654,
506,
667
],
"score": 1.0,
"content": "generalize our methodology to an other MI task of BCI Competition IV-2a dataset and believe that it",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
105,
665,
293,
678
],
"spans": [
{
"bbox": [
105,
665,
293,
678
],
"score": 1.0,
"content": "can be easily adapted to other BCI paradigms.",
"type": "text"
}
],
"index": 51
}
],
"index": 44.5
},
{
"type": "text",
"bbox": [
108,
682,
306,
693
],
"lines": [
{
"bbox": [
106,
681,
307,
694
],
"spans": [
{
"bbox": [
106,
681,
307,
694
],
"score": 1.0,
"content": "Main contributions. Our main contributions are:",
"type": "text"
}
],
"index": 52
}
],
"index": 52
},
{
"type": "text",
"bbox": [
134,
699,
504,
722
],
"lines": [
{
"bbox": [
132,
697,
505,
712
],
"spans": [
{
"bbox": [
132,
697,
505,
712
],
"score": 1.0,
"content": "• We design a new method to generate smooth adversarial perturbations that are",
"type": "text"
}
],
"index": 53
},
{
"bbox": [
141,
709,
392,
725
],
"spans": [
{
"bbox": [
141,
709,
392,
725
],
"score": 1.0,
"content": "physiologically plausible and imperceptible to the human eye.",
"type": "text"
}
],
"index": 54
}
],
"index": 53.5
}
],
"page_idx": 1,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
309,
39
],
"spans": [
{
"bbox": [
106,
25,
309,
39
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
752,
309,
760
],
"lines": [
{
"bbox": [
301,
750,
310,
763
],
"spans": [
{
"bbox": [
301,
750,
310,
763
],
"score": 1.0,
"content": "",
"type": "text",
"height": 13,
"width": 9
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
170
],
"lines": [],
"index": 3.5,
"bbox_fs": [
105,
83,
505,
171
],
"lines_deleted": true
},
{
"type": "text",
"bbox": [
107,
176,
505,
308
],
"lines": [
{
"bbox": [
106,
175,
506,
189
],
"spans": [
{
"bbox": [
106,
175,
506,
189
],
"score": 1.0,
"content": "Challenges: Designing natural attacks and modeling its propagation. Unlike in audio",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
187,
506,
200
],
"spans": [
{
"bbox": [
106,
187,
506,
200
],
"score": 1.0,
"content": "applications where the signal can simply propagate over-the-air and is sensed by a microphone, extra",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
198,
505,
210
],
"spans": [
{
"bbox": [
106,
198,
505,
210
],
"score": 1.0,
"content": "modeling is required to evaluate the signal propagation in BCIs based on the physical properties of the",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
208,
506,
223
],
"spans": [
{
"bbox": [
105,
208,
506,
223
],
"score": 1.0,
"content": "biological tissues. In this work, rather than assuming a “jamming” module between the preprocessing",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
220,
505,
232
],
"spans": [
{
"bbox": [
106,
220,
505,
232
],
"score": 1.0,
"content": "and the classification steps as in related works, we consider a more realistic and practically applicable",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
106,
231,
506,
243
],
"spans": [
{
"bbox": [
106,
231,
506,
243
],
"score": 1.0,
"content": "attack scenario where the adversarial perturbations are introduced at the source of the data acquisition,",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
243,
505,
254
],
"spans": [
{
"bbox": [
106,
243,
505,
254
],
"score": 1.0,
"content": "as showcased in Figure 6 in Appendix A. This can be achieved, for example, via electromagnetic",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
253,
505,
264
],
"spans": [
{
"bbox": [
105,
253,
505,
264
],
"score": 1.0,
"content": "waves delivered to the environment (Dutta, 2020) or via transcranial current stimulation with electrical",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
263,
506,
277
],
"spans": [
{
"bbox": [
105,
263,
506,
277
],
"score": 1.0,
"content": "current delivered directly to the scalp (Bodranghien et al., 2017; Fertonani et al., 2015), by exploiting",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
274,
506,
287
],
"spans": [
{
"bbox": [
105,
274,
506,
287
],
"score": 1.0,
"content": "wearable devices, such as smart glasses or over-ear headsets (Flowneuroscience, 2021; Marin et al.,",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
286,
506,
299
],
"spans": [
{
"bbox": [
106,
286,
506,
299
],
"score": 1.0,
"content": "2020). The adversarial perturbations translate into electrical signals propagating over the scalp and",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
297,
345,
309
],
"spans": [
{
"bbox": [
106,
297,
345,
309
],
"score": 1.0,
"content": "are sensed by the electrodes in addition to the EEG signals.",
"type": "text"
}
],
"index": 19
}
],
"index": 13.5,
"bbox_fs": [
105,
175,
506,
309
]
},
{
"type": "text",
"bbox": [
107,
313,
505,
390
],
"lines": [
{
"bbox": [
106,
313,
505,
326
],
"spans": [
{
"bbox": [
106,
313,
505,
326
],
"score": 1.0,
"content": "To guarantee the imperceptibility of the attacks, previous works in BCIs create perturbations that",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
325,
505,
336
],
"spans": [
{
"bbox": [
106,
325,
505,
336
],
"score": 1.0,
"content": "are small in amplitude (Zhang & Wu, 2019; Jiang et al., 2019; Liu et al., 2021), limiting the attack",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
106,
336,
506,
348
],
"spans": [
{
"bbox": [
106,
336,
506,
348
],
"score": 1.0,
"content": "success rate (ASR). Increased perturbation’s amplitude yields higher ASR (Meng et al., 2021), but",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
347,
506,
359
],
"spans": [
{
"bbox": [
106,
347,
506,
359
],
"score": 1.0,
"content": "makes the attack more easily detectable. Moreover, the generated perturbations are square-shaped,",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
358,
505,
370
],
"spans": [
{
"bbox": [
106,
358,
505,
370
],
"score": 1.0,
"content": "which is implausible for biosignals. Han et al. (2020) are the first to observe square-wave artifacts",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
367,
506,
382
],
"spans": [
{
"bbox": [
105,
367,
506,
382
],
"score": 1.0,
"content": "in biosignals’ attacks and propose smooth perturbations for electrocardiograms (ECGs). No similar",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
380,
243,
391
],
"spans": [
{
"bbox": [
106,
380,
243,
391
],
"score": 1.0,
"content": "works have been found for EEGs.",
"type": "text"
}
],
"index": 26
}
],
"index": 23,
"bbox_fs": [
105,
313,
506,
391
]
},
{
"type": "text",
"bbox": [
106,
396,
505,
517
],
"lines": [
{
"bbox": [
105,
395,
505,
409
],
"spans": [
{
"bbox": [
105,
395,
505,
409
],
"score": 1.0,
"content": "This work: Practical attacks on BCI models. To address the above technical challenges, and",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
407,
505,
420
],
"spans": [
{
"bbox": [
106,
407,
505,
420
],
"score": 1.0,
"content": "for analyzing the vulnerability of embedded BCI models in practical scenarios, we design a new",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
419,
505,
431
],
"spans": [
{
"bbox": [
106,
419,
505,
431
],
"score": 1.0,
"content": "attack algorithm that generates smooth adversarial examples based on the signals’ first derivative and",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
430,
504,
441
],
"spans": [
{
"bbox": [
106,
430,
504,
441
],
"score": 1.0,
"content": "model its propagation over the scalp based on a realistic head model by taking into consideration the",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
441,
505,
453
],
"spans": [
{
"bbox": [
106,
441,
505,
453
],
"score": 1.0,
"content": "attack source and the electrical and physical properties of the conducting tissues. This enables the",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
451,
506,
464
],
"spans": [
{
"bbox": [
105,
451,
506,
464
],
"score": 1.0,
"content": "creation of practically effective perturbations, that can be delivered by an external device to attack",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
461,
506,
474
],
"spans": [
{
"bbox": [
105,
461,
506,
474
],
"score": 1.0,
"content": "EEG-based BCIs at the source of signal acquisition. We attack the most energy-efficient network that",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
472,
506,
486
],
"spans": [
{
"bbox": [
105,
472,
506,
486
],
"score": 1.0,
"content": "has been embedded on microcontrollers for smart wearable BCIs called EEGNet (Lawhern et al.,",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
484,
504,
496
],
"spans": [
{
"bbox": [
106,
484,
504,
496
],
"score": 1.0,
"content": "2018; Schneider et al., 2020). It is a resource-friendly convolutional neural network (CNN) and is the",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
495,
506,
509
],
"spans": [
{
"bbox": [
105,
495,
506,
509
],
"score": 1.0,
"content": "SoA in terms of accuracy and energy-efficiency trade-off (Belwafi et al., 2018; Malekmohammadi",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
507,
325,
518
],
"spans": [
{
"bbox": [
106,
507,
325,
518
],
"score": 1.0,
"content": "et al., 2019; Wang et al., 2020; Schneider et al., 2020).",
"type": "text"
}
],
"index": 37
}
],
"index": 32,
"bbox_fs": [
105,
395,
506,
518
]
},
{
"type": "text",
"bbox": [
106,
523,
506,
676
],
"lines": [
{
"bbox": [
105,
522,
506,
535
],
"spans": [
{
"bbox": [
105,
522,
506,
535
],
"score": 1.0,
"content": "We evaluate our methods and show experimental results on BCIs based on the motor imagery (MI)",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
534,
506,
547
],
"spans": [
{
"bbox": [
105,
534,
506,
547
],
"score": 1.0,
"content": "paradigm, which is of special interest among others because it can be asynchronously self-paced",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
104,
543,
507,
559
],
"spans": [
{
"bbox": [
104,
543,
507,
559
],
"score": 1.0,
"content": "without external stimuli (Freer & Yang, 2020). By imagining the movement of different body parts,",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
555,
506,
569
],
"spans": [
{
"bbox": [
105,
555,
506,
569
],
"score": 1.0,
"content": "the decoded intention is translated into control signals. It is widely applied in several BCI applications,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
567,
506,
579
],
"spans": [
{
"bbox": [
105,
567,
506,
579
],
"score": 1.0,
"content": "such as the control of wheelchairss (Yu et al., 2018), prosthetic armss (Elstob & Secco, 2016), ground",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
578,
506,
590
],
"spans": [
{
"bbox": [
105,
578,
506,
590
],
"score": 1.0,
"content": "vehicles (Zhuang et al., 2021), and in communication (Brumberg et al., 2016). It has been proven",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
588,
506,
601
],
"spans": [
{
"bbox": [
105,
588,
506,
601
],
"score": 1.0,
"content": "to be the most difficult task to be attacked among the most common BCI paradigms (Zhang & Wu,",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
599,
506,
612
],
"spans": [
{
"bbox": [
105,
599,
506,
612
],
"score": 1.0,
"content": "2019; Meng et al., 2021). We evaluate our methods by “fooling” the victim model to always predict",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
610,
506,
623
],
"spans": [
{
"bbox": [
105,
610,
506,
623
],
"score": 1.0,
"content": "“rest” class. This essentially yields a DoS attack, because resting-state EEG signals are generally",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
621,
506,
633
],
"spans": [
{
"bbox": [
105,
621,
506,
633
],
"score": 1.0,
"content": "interpreted as no subject’s intention decoded, i.e., no control action needs to be taken by the BCI",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
632,
506,
645
],
"spans": [
{
"bbox": [
105,
632,
506,
645
],
"score": 1.0,
"content": "system (Yu et al., 2018). While for healthy subjects it might solely cause user frustration and financial",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
106,
643,
505,
655
],
"spans": [
{
"bbox": [
106,
643,
505,
655
],
"score": 1.0,
"content": "losses, for severely paralyzed patients it can lead to loss of communication and independence. We",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
654,
506,
667
],
"spans": [
{
"bbox": [
105,
654,
506,
667
],
"score": 1.0,
"content": "generalize our methodology to an other MI task of BCI Competition IV-2a dataset and believe that it",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
105,
665,
293,
678
],
"spans": [
{
"bbox": [
105,
665,
293,
678
],
"score": 1.0,
"content": "can be easily adapted to other BCI paradigms.",
"type": "text"
}
],
"index": 51
}
],
"index": 44.5,
"bbox_fs": [
104,
522,
507,
678
]
},
{
"type": "text",
"bbox": [
108,
682,
306,
693
],
"lines": [
{
"bbox": [
106,
681,
307,
694
],
"spans": [
{
"bbox": [
106,
681,
307,
694
],
"score": 1.0,
"content": "Main contributions. Our main contributions are:",
"type": "text"
}
],
"index": 52
}
],
"index": 52,
"bbox_fs": [
106,
681,
307,
694
]
},
{
"type": "text",
"bbox": [
134,
699,
504,
722
],
"lines": [
{
"bbox": [
132,
697,
505,
712
],
"spans": [
{
"bbox": [
132,
697,
505,
712
],
"score": 1.0,
"content": "• We design a new method to generate smooth adversarial perturbations that are",
"type": "text"
}
],
"index": 53
},
{
"bbox": [
141,
709,
392,
725
],
"spans": [
{
"bbox": [
141,
709,
392,
725
],
"score": 1.0,
"content": "physiologically plausible and imperceptible to the human eye.",
"type": "text"
}
],
"index": 54
}
],
"index": 53.5,
"bbox_fs": [
132,
697,
505,
725
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
133,
82,
505,
152
],
"lines": [
{
"bbox": [
133,
82,
506,
95
],
"spans": [
{
"bbox": [
133,
82,
506,
95
],
"score": 1.0,
"content": "• We consider a practical scenario where the perturbation is added at the signal acquisition",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
140,
93,
506,
107
],
"spans": [
{
"bbox": [
140,
93,
506,
107
],
"score": 1.0,
"content": "source and model its propagation constrained by the physical properties of the human scalp.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
132,
107,
506,
120
],
"spans": [
{
"bbox": [
132,
107,
506,
120
],
"score": 1.0,
"content": "• The first study of adversarial perturbations in BCI to consider the practical scenario of",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
141,
119,
506,
130
],
"spans": [
{
"bbox": [
141,
119,
506,
130
],
"score": 1.0,
"content": "smart edge computing and physical signal propagation. We create both local and global",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
141,
129,
506,
141
],
"spans": [
{
"bbox": [
141,
129,
461,
141
],
"score": 1.0,
"content": "perturbations and show that our attacks consistently achieve a success rate of",
"type": "text"
},
{
"bbox": [
462,
129,
493,
140
],
"score": 0.87,
"content": "> 5 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
493,
129,
506,
141
],
"score": 1.0,
"content": "in",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
142,
141,
498,
153
],
"spans": [
{
"bbox": [
142,
141,
498,
153
],
"score": 1.0,
"content": "different settings pointing to the significant vulnerability of the SoA embedded EEGNet.",
"type": "text"
}
],
"index": 5
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
107,
157,
505,
180
],
"lines": [
{
"bbox": [
105,
157,
505,
169
],
"spans": [
{
"bbox": [
105,
157,
505,
169
],
"score": 1.0,
"content": "We hope that our work raises awareness for potential risks and motivates the future development of",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
169,
226,
180
],
"spans": [
{
"bbox": [
105,
169,
226,
180
],
"score": 1.0,
"content": "appropriate countermeasures.",
"type": "text"
}
],
"index": 7
}
],
"index": 6.5
},
{
"type": "title",
"bbox": [
108,
194,
201,
208
],
"lines": [
{
"bbox": [
104,
193,
201,
210
],
"spans": [
{
"bbox": [
104,
193,
201,
210
],
"score": 1.0,
"content": "2 BACKGROUND",
"type": "text"
}
],
"index": 8
}
],
"index": 8
},
{
"type": "title",
"bbox": [
108,
218,
241,
231
],
"lines": [
{
"bbox": [
105,
218,
241,
232
],
"spans": [
{
"bbox": [
105,
218,
241,
232
],
"score": 1.0,
"content": "2.1 CLASSIFICATION IN BCIS",
"type": "text"
}
],
"index": 9
}
],
"index": 9
},
{
"type": "text",
"bbox": [
106,
239,
505,
286
],
"lines": [
{
"bbox": [
105,
238,
505,
253
],
"spans": [
{
"bbox": [
105,
238,
505,
253
],
"score": 1.0,
"content": "We first describe the commonly used approach in BCIs for classification, consisting of a preprocessing",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
250,
505,
263
],
"spans": [
{
"bbox": [
106,
250,
449,
263
],
"score": 1.0,
"content": "step and a classifier. The brain activity is recorded with an EEG device which samples",
"type": "text"
},
{
"bbox": [
449,
251,
467,
262
],
"score": 0.9,
"content": "N _ { c h }",
"type": "inline_equation"
},
{
"bbox": [
467,
250,
505,
263
],
"score": 1.0,
"content": "channels",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
261,
506,
275
],
"spans": [
{
"bbox": [
105,
261,
133,
275
],
"score": 1.0,
"content": "at rate",
"type": "text"
},
{
"bbox": [
134,
263,
145,
274
],
"score": 0.87,
"content": "F _ { s }",
"type": "inline_equation"
},
{
"bbox": [
146,
261,
228,
275
],
"score": 1.0,
"content": ". We define one trial",
"type": "text"
},
{
"bbox": [
229,
263,
235,
274
],
"score": 0.8,
"content": "j",
"type": "inline_equation"
},
{
"bbox": [
235,
261,
247,
275
],
"score": 1.0,
"content": "as",
"type": "text"
},
{
"bbox": [
248,
262,
295,
275
],
"score": 0.92,
"content": "( \\mathbf { X } ^ { ( j ) } , y ^ { ( j ) } )",
"type": "inline_equation"
},
{
"bbox": [
295,
261,
326,
275
],
"score": 1.0,
"content": ", where",
"type": "text"
},
{
"bbox": [
326,
261,
429,
275
],
"score": 0.92,
"content": "y ^ { ( j ) } \\in \\{ 0 , 1 , . . . , N _ { c l } - 1 \\}",
"type": "inline_equation"
},
{
"bbox": [
429,
261,
506,
275
],
"score": 1.0,
"content": "is the true label of",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
107,
273,
409,
288
],
"spans": [
{
"bbox": [
107,
275,
122,
286
],
"score": 0.87,
"content": "N _ { c l }",
"type": "inline_equation"
},
{
"bbox": [
123,
273,
180,
288
],
"score": 1.0,
"content": "MI tasks, and",
"type": "text"
},
{
"bbox": [
180,
274,
250,
285
],
"score": 0.92,
"content": "\\mathbf { X } ^ { ( j ) } \\in \\mathbb { R } ^ { N _ { s } \\times N _ { c h } }",
"type": "inline_equation"
},
{
"bbox": [
250,
273,
409,
288
],
"score": 1.0,
"content": "the multi-channel recording defined as",
"type": "text"
}
],
"index": 13
}
],
"index": 11.5
},
{
"type": "interline_equation",
"bbox": [
236,
287,
374,
308
],
"lines": [
{
"bbox": [
236,
287,
374,
308
],
"spans": [
{
"bbox": [
236,
287,
374,
308
],
"score": 0.94,
"content": "\\mathbf { X } ^ { ( j ) } : = \\left( \\mathbf { x } _ { 0 } ^ { ( j ) } , \\mathbf { x } _ { 1 } ^ { ( j ) } , . . . , \\mathbf { x } _ { N _ { c h } - 1 } ^ { ( j ) } \\right) ,",
"type": "interline_equation",
"image_path": "4adef2bd32024f0155b08bd021946820057f4e0a6d08b9b081db9e63f5cb922a.jpg"
}
]
}
],
"index": 14,
"virtual_lines": [
{
"bbox": [
236,
287,
374,
308
],
"spans": [],
"index": 14
}
]
},
{
"type": "text",
"bbox": [
109,
312,
505,
336
],
"lines": [
{
"bbox": [
107,
309,
504,
325
],
"spans": [
{
"bbox": [
107,
309,
126,
325
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
127,
309,
173,
325
],
"score": 0.93,
"content": "\\mathbf { x } _ { i } ^ { ( j ) } \\in \\mathbb { R } ^ { N _ { s } }",
"type": "inline_equation"
},
{
"bbox": [
174,
309,
326,
325
],
"score": 1.0,
"content": "corresponding to the recording of the",
"type": "text"
},
{
"bbox": [
327,
313,
332,
324
],
"score": 0.82,
"content": "j",
"type": "inline_equation"
},
{
"bbox": [
333,
309,
394,
325
],
"score": 1.0,
"content": "-th trial and the",
"type": "text"
},
{
"bbox": [
395,
313,
399,
322
],
"score": 0.78,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
400,
309,
490,
325
],
"score": 1.0,
"content": "-th channel containing",
"type": "text"
},
{
"bbox": [
491,
313,
504,
323
],
"score": 0.85,
"content": "N _ { s }",
"type": "inline_equation"
}
],
"index": 15
},
{
"bbox": [
107,
323,
390,
337
],
"spans": [
{
"bbox": [
107,
323,
285,
337
],
"score": 1.0,
"content": "temporal samples. For simplicity, we denote",
"type": "text"
},
{
"bbox": [
286,
324,
330,
335
],
"score": 0.91,
"content": "\\mathbf { X } : = \\mathbf { X } ^ { ( j ) }",
"type": "inline_equation"
},
{
"bbox": [
330,
323,
349,
337
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
349,
324,
387,
336
],
"score": 0.93,
"content": "y : = y ^ { ( j ) }",
"type": "inline_equation"
},
{
"bbox": [
387,
323,
390,
337
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 16
}
],
"index": 15.5
},
{
"type": "text",
"bbox": [
106,
341,
503,
364
],
"lines": [
{
"bbox": [
106,
341,
505,
354
],
"spans": [
{
"bbox": [
106,
341,
505,
354
],
"score": 1.0,
"content": "The EEG recordings are often preprocessed with a band-pass filter, e.g., using a Fast Fourier Transform",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
351,
344,
366
],
"spans": [
{
"bbox": [
105,
351,
154,
366
],
"score": 1.0,
"content": "(FFT) filter",
"type": "text"
},
{
"bbox": [
154,
353,
179,
365
],
"score": 0.92,
"content": "h _ { b p } ( \\cdot )",
"type": "inline_equation"
},
{
"bbox": [
179,
351,
344,
366
],
"score": 1.0,
"content": ", before being fed to a classifier, yielding",
"type": "text"
}
],
"index": 18
}
],
"index": 17.5
},
{
"type": "interline_equation",
"bbox": [
191,
365,
420,
379
],
"lines": [
{
"bbox": [
191,
365,
420,
379
],
"spans": [
{
"bbox": [
191,
365,
420,
379
],
"score": 0.88,
"content": "\\mathbf { X } _ { b p } = H _ { b p } ( \\mathbf { X } ) = \\left( h _ { b p } ( \\mathbf { x } _ { 0 } ) , h _ { b p } ( \\mathbf { x } _ { 1 } ) , . . . , h _ { b p } ( \\mathbf { x } _ { N _ { c h } - 1 } ) \\right) .",
"type": "interline_equation",
"image_path": "92578742fa44b8eb625a3c6e4e087ff7ad891f1315813fccd22129d3165cbf68.jpg"
}
]
}
],
"index": 19,
"virtual_lines": [
{
"bbox": [
191,
365,
420,
379
],
"spans": [],
"index": 19
}
]
},
{
"type": "text",
"bbox": [
107,
380,
506,
414
],
"lines": [
{
"bbox": [
105,
378,
505,
393
],
"spans": [
{
"bbox": [
105,
378,
237,
393
],
"score": 1.0,
"content": "Finally, the preprocessed signal",
"type": "text"
},
{
"bbox": [
238,
380,
256,
392
],
"score": 0.9,
"content": "\\mathbf { X } _ { b p }",
"type": "inline_equation"
},
{
"bbox": [
256,
378,
401,
393
],
"score": 1.0,
"content": "is classified with a trainable model",
"type": "text"
},
{
"bbox": [
402,
380,
409,
391
],
"score": 0.84,
"content": "f",
"type": "inline_equation"
},
{
"bbox": [
409,
378,
482,
393
],
"score": 1.0,
"content": "and is mapped to",
"type": "text"
},
{
"bbox": [
482,
381,
505,
392
],
"score": 0.82,
"content": "\\mathbf { p } : =",
"type": "inline_equation"
}
],
"index": 20
},
{
"bbox": [
107,
390,
506,
405
],
"spans": [
{
"bbox": [
107,
391,
142,
404
],
"score": 0.89,
"content": "f \\left( { \\bf { X } } _ { b p } \\right)",
"type": "inline_equation"
},
{
"bbox": [
142,
390,
171,
405
],
"score": 1.0,
"content": ", where",
"type": "text"
},
{
"bbox": [
171,
391,
210,
403
],
"score": 0.91,
"content": "\\mathbf { p } \\in \\mathbb { R } ^ { N _ { c l } }",
"type": "inline_equation"
},
{
"bbox": [
210,
390,
506,
405
],
"score": 1.0,
"content": "contains the output probabilities, e.g., originating from a softmax activation",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
402,
494,
416
],
"spans": [
{
"bbox": [
105,
402,
187,
416
],
"score": 1.0,
"content": "as final operation in",
"type": "text"
},
{
"bbox": [
188,
403,
195,
414
],
"score": 0.84,
"content": "f",
"type": "inline_equation"
},
{
"bbox": [
195,
402,
314,
416
],
"score": 1.0,
"content": ". The model’s final prediction",
"type": "text"
},
{
"bbox": [
315,
403,
321,
414
],
"score": 0.85,
"content": "\\hat { y }",
"type": "inline_equation"
},
{
"bbox": [
321,
402,
482,
416
],
"score": 1.0,
"content": "is the index with the maximum score in",
"type": "text"
},
{
"bbox": [
482,
404,
489,
414
],
"score": 0.33,
"content": "\\mathbf { p }",
"type": "inline_equation"
},
{
"bbox": [
490,
402,
494,
416
],
"score": 1.0,
"content": ":",
"type": "text"
}
],
"index": 22
}
],
"index": 21
},
{
"type": "interline_equation",
"bbox": [
219,
415,
391,
439
],
"lines": [
{
"bbox": [
219,
415,
391,
439
],
"spans": [
{
"bbox": [
219,
415,
391,
439
],
"score": 0.93,
"content": "\\boldsymbol { \\hat { y } } = \\boldsymbol { \\hat { f } } \\left( \\mathbf { X } _ { b p } \\right) = \\underset { y \\in \\{ 0 , \\ldots , N _ { c l } - 1 \\} } { \\mathrm { a r g m a x } } ~ f \\left( \\mathbf { X } _ { b p } \\right) [ y ] .",
"type": "interline_equation",
"image_path": "090de2464ceebb19449b8aaa8df74dfb9fcbf1932b76ec9c09e0616b0fd63a0c.jpg"
}
]
}
],
"index": 23,
"virtual_lines": [
{
"bbox": [
219,
415,
391,
439
],
"spans": [],
"index": 23
}
]
},
{
"type": "title",
"bbox": [
108,
450,
248,
461
],
"lines": [
{
"bbox": [
106,
450,
249,
462
],
"spans": [
{
"bbox": [
106,
450,
249,
462
],
"score": 1.0,
"content": "2.2 INSTANCE-BASED ATTACKS",
"type": "text"
}
],
"index": 24
}
],
"index": 24
},
{
"type": "text",
"bbox": [
106,
470,
505,
516
],
"lines": [
{
"bbox": [
106,
470,
505,
483
],
"spans": [
{
"bbox": [
106,
470,
315,
483
],
"score": 1.0,
"content": "Instance-based attacks try to fool an EEG classifier",
"type": "text"
},
{
"bbox": [
315,
471,
322,
482
],
"score": 0.86,
"content": "f",
"type": "inline_equation"
},
{
"bbox": [
322,
470,
442,
483
],
"score": 1.0,
"content": "to misclassify an EEG signal",
"type": "text"
},
{
"bbox": [
442,
471,
452,
481
],
"score": 0.58,
"content": "\\mathbf { X }",
"type": "inline_equation"
},
{
"bbox": [
452,
470,
505,
483
],
"score": 1.0,
"content": "to a targeted",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
481,
505,
493
],
"spans": [
{
"bbox": [
106,
481,
128,
493
],
"score": 1.0,
"content": "class",
"type": "text"
},
{
"bbox": [
129,
483,
138,
493
],
"score": 0.83,
"content": "y _ { t }",
"type": "inline_equation"
},
{
"bbox": [
138,
481,
398,
493
],
"score": 1.0,
"content": ". In this section, we describe the attack directly on the classifier",
"type": "text"
},
{
"bbox": [
398,
482,
405,
493
],
"score": 0.85,
"content": "f",
"type": "inline_equation"
},
{
"bbox": [
406,
481,
505,
493
],
"score": 1.0,
"content": "without considering the",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
493,
506,
505
],
"spans": [
{
"bbox": [
105,
493,
166,
505
],
"score": 1.0,
"content": "preprocessing",
"type": "text"
},
{
"bbox": [
166,
493,
183,
505
],
"score": 0.91,
"content": "H _ { b p }",
"type": "inline_equation"
},
{
"bbox": [
184,
493,
506,
505
],
"score": 1.0,
"content": "; the inclusion of the preprocessing is described in Section 3.3. We define an",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
104,
503,
365,
517
],
"spans": [
{
"bbox": [
104,
503,
216,
517
],
"score": 1.0,
"content": "adversarial example as any",
"type": "text"
},
{
"bbox": [
216,
504,
324,
515
],
"score": 0.89,
"content": "\\mathbf { X } ^ { * } = \\mathbf { X } + \\mathbf { V } \\in \\mathbb { R } ^ { N _ { s } \\times N _ { c h } }",
"type": "inline_equation"
},
{
"bbox": [
324,
503,
365,
517
],
"score": 1.0,
"content": "such that",
"type": "text"
}
],
"index": 28
}
],
"index": 26.5
},
{
"type": "interline_equation",
"bbox": [
259,
517,
351,
533
],
"lines": [
{
"bbox": [
259,
517,
351,
533
],
"spans": [
{
"bbox": [
259,
517,
351,
533
],
"score": 0.88,
"content": "\\hat { f } \\left( \\mathbf { X } \\right) \\neq \\hat { f } \\left( \\mathbf { X } ^ { * } \\right) = y _ { t } .",
"type": "interline_equation",
"image_path": "ae96fbca25561b728f407743e94ef7daae4c068465648eca1989b79f54f18f8f.jpg"
}
]
}
],
"index": 29,
"virtual_lines": [
{
"bbox": [
259,
517,
351,
533
],
"spans": [],
"index": 29
}
]
},
{
"type": "text",
"bbox": [
106,
541,
503,
563
],
"lines": [
{
"bbox": [
104,
539,
503,
554
],
"spans": [
{
"bbox": [
104,
539,
443,
554
],
"score": 1.0,
"content": "FGSM. The FGSM (Goodfellow et al., 2015) generates an adversarial perturbation",
"type": "text"
},
{
"bbox": [
443,
540,
503,
551
],
"score": 0.9,
"content": "\\mathbf { V } \\in \\mathbb { R } ^ { N _ { s } \\times N _ { c h } }",
"type": "inline_equation"
}
],
"index": 30
},
{
"bbox": [
105,
551,
437,
565
],
"spans": [
{
"bbox": [
105,
551,
162,
565
],
"score": 1.0,
"content": "of magnitude",
"type": "text"
},
{
"bbox": [
162,
554,
167,
562
],
"score": 0.67,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
168,
551,
437,
565
],
"score": 1.0,
"content": "which points in the negative direction of a loss function’s gradient:",
"type": "text"
}
],
"index": 31
}
],
"index": 30.5
},
{
"type": "interline_equation",
"bbox": [
239,
564,
370,
578
],
"lines": [
{
"bbox": [
239,
564,
370,
578
],
"spans": [
{
"bbox": [
239,
564,
370,
578
],
"score": 0.9,
"content": "\\mathbf { V } = - \\epsilon \\cdot \\mathrm { s i g n } \\left( \\nabla _ { \\mathbf { X } } \\cdot l \\left( \\mathbf { X } , y _ { t } \\right) \\right) ,",
"type": "interline_equation",
"image_path": "0d4290c3afb9da2ee94681d4b078c69233401f350f53f236fd055d34090accc5.jpg"
}
]
}
],
"index": 32,
"virtual_lines": [
{
"bbox": [
239,
564,
370,
578
],
"spans": [],
"index": 32
}
]
},
{
"type": "text",
"bbox": [
106,
578,
347,
590
],
"lines": [
{
"bbox": [
105,
577,
345,
592
],
"spans": [
{
"bbox": [
105,
577,
345,
592
],
"score": 1.0,
"content": "where the loss function contains the negative log likelihood",
"type": "text"
}
],
"index": 33
}
],
"index": 33
},
{
"type": "interline_equation",
"bbox": [
212,
591,
398,
604
],
"lines": [
{
"bbox": [
212,
591,
398,
604
],
"spans": [
{
"bbox": [
212,
591,
398,
604
],
"score": 0.9,
"content": "l ( \\mathbf { X } , y _ { t } ) = - \\log \\left( { \\mathbf { p } } [ y _ { t } ] \\right) = - \\log \\left( f \\left( \\mathbf { X } \\right) [ y _ { t } ] \\right) .",
"type": "interline_equation",
"image_path": "76d960042d0c715217ff55097d0087d7852aed951636159758fb7c8f7d3fa671.jpg"
}
]
}
],
"index": 34,
"virtual_lines": [
{
"bbox": [
212,
591,
398,
604
],
"spans": [],
"index": 34
}
]
},
{
"type": "text",
"bbox": [
105,
606,
504,
628
],
"lines": [
{
"bbox": [
105,
605,
505,
619
],
"spans": [
{
"bbox": [
105,
605,
119,
619
],
"score": 1.0,
"content": "As",
"type": "text"
},
{
"bbox": [
119,
607,
127,
617
],
"score": 0.37,
"content": "\\mathbf { p }",
"type": "inline_equation"
},
{
"bbox": [
128,
605,
505,
619
],
"score": 1.0,
"content": "is the output of the softmax activation function, equation 6 becomes a cross-entropy loss which",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
616,
382,
629
],
"spans": [
{
"bbox": [
105,
616,
194,
629
],
"score": 1.0,
"content": "maximizes the output",
"type": "text"
},
{
"bbox": [
195,
616,
216,
628
],
"score": 0.91,
"content": "\\mathbf { p } [ y _ { t } ]",
"type": "inline_equation"
},
{
"bbox": [
217,
616,
382,
629
],
"score": 1.0,
"content": "while minimizing the remaining outputs.",
"type": "text"
}
],
"index": 36
}
],
"index": 35.5
},
{
"type": "text",
"bbox": [
106,
638,
505,
694
],
"lines": [
{
"bbox": [
105,
638,
506,
652
],
"spans": [
{
"bbox": [
105,
638,
506,
652
],
"score": 1.0,
"content": "PGD. The projected gradient descent (PGD) (Madry et al., 2018) is a variant of the basic iterative",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
650,
506,
662
],
"spans": [
{
"bbox": [
106,
650,
506,
662
],
"score": 1.0,
"content": "method (Kurakin et al.), generally considered to be more effective than FGSM. PGD aims to find a",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
660,
506,
675
],
"spans": [
{
"bbox": [
105,
660,
316,
675
],
"score": 1.0,
"content": "perturbation by iteratively taking small steps of size",
"type": "text"
},
{
"bbox": [
316,
663,
324,
671
],
"score": 0.79,
"content": "\\alpha",
"type": "inline_equation"
},
{
"bbox": [
325,
660,
506,
675
],
"score": 1.0,
"content": "in the gradient’s direction and projecting the",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
672,
505,
684
],
"spans": [
{
"bbox": [
106,
672,
505,
684
],
"score": 1.0,
"content": "resulting perturbation back to the sample’s neighborhood after each iteration. We randomly initialize",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
682,
472,
696
],
"spans": [
{
"bbox": [
106,
682,
188,
696
],
"score": 1.0,
"content": "the attack inside the",
"type": "text"
},
{
"bbox": [
189,
684,
205,
694
],
"score": 0.9,
"content": "L _ { \\infty }",
"type": "inline_equation"
},
{
"bbox": [
205,
682,
261,
696
],
"score": 1.0,
"content": "ball of radius",
"type": "text"
},
{
"bbox": [
261,
685,
267,
693
],
"score": 0.73,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
267,
682,
354,
696
],
"score": 1.0,
"content": "and update the attack",
"type": "text"
},
{
"bbox": [
355,
683,
378,
695
],
"score": 0.91,
"content": "\\mathbf { V } _ { t + 1 }",
"type": "inline_equation"
},
{
"bbox": [
378,
682,
446,
696
],
"score": 1.0,
"content": "for any iteration",
"type": "text"
},
{
"bbox": [
446,
684,
451,
693
],
"score": 0.76,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
451,
682,
472,
696
],
"score": 1.0,
"content": "with",
"type": "text"
}
],
"index": 41
}
],
"index": 39
},
{
"type": "interline_equation",
"bbox": [
200,
696,
408,
709
],
"lines": [
{
"bbox": [
200,
696,
408,
709
],
"spans": [
{
"bbox": [
200,
696,
408,
709
],
"score": 0.83,
"content": "\\mathbf { V } _ { t + 1 } = \\mathrm { c l i p } _ { \\epsilon } \\left( \\mathbf { V } _ { t } - \\alpha \\cdot \\mathrm { s i g n } \\left( \\nabla _ { \\mathbf { V } } l \\left( \\mathbf { X } + \\mathbf { V } _ { t } , y _ { t } \\right) \\right) \\right) ,",
"type": "interline_equation",
"image_path": "5018cbeb47f39276c0e84f65e695e567cca2809727b109293d37f372e63a064d.jpg"
}
]
}
],
"index": 42,
"virtual_lines": [
{
"bbox": [
200,
696,
408,
709
],
"spans": [],
"index": 42
}
]
},
{
"type": "text",
"bbox": [
106,
710,
507,
732
],
"lines": [
{
"bbox": [
106,
709,
504,
721
],
"spans": [
{
"bbox": [
106,
709,
134,
721
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
712,
142,
720
],
"score": 0.78,
"content": "\\alpha",
"type": "inline_equation"
},
{
"bbox": [
142,
709,
254,
721
],
"score": 1.0,
"content": "is a step size smaller than",
"type": "text"
},
{
"bbox": [
254,
712,
260,
720
],
"score": 0.68,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
260,
709,
504,
721
],
"score": 1.0,
"content": "which decays linearly with each iteration and the function",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
720,
353,
734
],
"spans": [
{
"bbox": [
106,
720,
139,
733
],
"score": 0.87,
"content": "\\mathrm { c l i p } _ { \\epsilon } \\left( \\cdot \\right)",
"type": "inline_equation"
},
{
"bbox": [
139,
720,
343,
734
],
"score": 1.0,
"content": "clips the signal at the maximum desired amplitude",
"type": "text"
},
{
"bbox": [
344,
723,
349,
730
],
"score": 0.72,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
349,
720,
353,
734
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 44
}
],
"index": 43.5
}
],
"page_idx": 2,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
309,
760
],
"lines": [
{
"bbox": [
301,
750,
310,
762
],
"spans": [
{
"bbox": [
301,
750,
310,
762
],
"score": 1.0,
"content": "3",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
133,
82,
505,
152
],
"lines": [
{
"bbox": [
133,
82,
506,
95
],
"spans": [
{
"bbox": [
133,
82,
506,
95
],
"score": 1.0,
"content": "• We consider a practical scenario where the perturbation is added at the signal acquisition",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
140,
93,
506,
107
],
"spans": [
{
"bbox": [
140,
93,
506,
107
],
"score": 1.0,
"content": "source and model its propagation constrained by the physical properties of the human scalp.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
132,
107,
506,
120
],
"spans": [
{
"bbox": [
132,
107,
506,
120
],
"score": 1.0,
"content": "• The first study of adversarial perturbations in BCI to consider the practical scenario of",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
141,
119,
506,
130
],
"spans": [
{
"bbox": [
141,
119,
506,
130
],
"score": 1.0,
"content": "smart edge computing and physical signal propagation. We create both local and global",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
141,
129,
506,
141
],
"spans": [
{
"bbox": [
141,
129,
461,
141
],
"score": 1.0,
"content": "perturbations and show that our attacks consistently achieve a success rate of",
"type": "text"
},
{
"bbox": [
462,
129,
493,
140
],
"score": 0.87,
"content": "> 5 0 \\%",
"type": "inline_equation"
},
{
"bbox": [
493,
129,
506,
141
],
"score": 1.0,
"content": "in",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
142,
141,
498,
153
],
"spans": [
{
"bbox": [
142,
141,
498,
153
],
"score": 1.0,
"content": "different settings pointing to the significant vulnerability of the SoA embedded EEGNet.",
"type": "text"
}
],
"index": 5
}
],
"index": 2.5,
"bbox_fs": [
132,
82,
506,
153
]
},
{
"type": "text",
"bbox": [
107,
157,
505,
180
],
"lines": [
{
"bbox": [
105,
157,
505,
169
],
"spans": [
{
"bbox": [
105,
157,
505,
169
],
"score": 1.0,
"content": "We hope that our work raises awareness for potential risks and motivates the future development of",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
169,
226,
180
],
"spans": [
{
"bbox": [
105,
169,
226,
180
],
"score": 1.0,
"content": "appropriate countermeasures.",
"type": "text"
}
],
"index": 7
}
],
"index": 6.5,
"bbox_fs": [
105,
157,
505,
180
]
},
{
"type": "title",
"bbox": [
108,
194,
201,
208
],
"lines": [
{
"bbox": [
104,
193,
201,
210
],
"spans": [
{
"bbox": [
104,
193,
201,
210
],
"score": 1.0,
"content": "2 BACKGROUND",
"type": "text"
}
],
"index": 8
}
],
"index": 8
},
{
"type": "title",
"bbox": [
108,
218,
241,
231
],
"lines": [
{
"bbox": [
105,
218,
241,
232
],
"spans": [
{
"bbox": [
105,
218,
241,
232
],
"score": 1.0,
"content": "2.1 CLASSIFICATION IN BCIS",
"type": "text"
}
],
"index": 9
}
],
"index": 9
},
{
"type": "text",
"bbox": [
106,
239,
505,
286
],
"lines": [
{
"bbox": [
105,
238,
505,
253
],
"spans": [
{
"bbox": [
105,
238,
505,
253
],
"score": 1.0,
"content": "We first describe the commonly used approach in BCIs for classification, consisting of a preprocessing",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
250,
505,
263
],
"spans": [
{
"bbox": [
106,
250,
449,
263
],
"score": 1.0,
"content": "step and a classifier. The brain activity is recorded with an EEG device which samples",
"type": "text"
},
{
"bbox": [
449,
251,
467,
262
],
"score": 0.9,
"content": "N _ { c h }",
"type": "inline_equation"
},
{
"bbox": [
467,
250,
505,
263
],
"score": 1.0,
"content": "channels",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
261,
506,
275
],
"spans": [
{
"bbox": [
105,
261,
133,
275
],
"score": 1.0,
"content": "at rate",
"type": "text"
},
{
"bbox": [
134,
263,
145,
274
],
"score": 0.87,
"content": "F _ { s }",
"type": "inline_equation"
},
{
"bbox": [
146,
261,
228,
275
],
"score": 1.0,
"content": ". We define one trial",
"type": "text"
},
{
"bbox": [
229,
263,
235,
274
],
"score": 0.8,
"content": "j",
"type": "inline_equation"
},
{
"bbox": [
235,
261,
247,
275
],
"score": 1.0,
"content": "as",
"type": "text"
},
{
"bbox": [
248,
262,
295,
275
],
"score": 0.92,
"content": "( \\mathbf { X } ^ { ( j ) } , y ^ { ( j ) } )",
"type": "inline_equation"
},
{
"bbox": [
295,
261,
326,
275
],
"score": 1.0,
"content": ", where",
"type": "text"
},
{
"bbox": [
326,
261,
429,
275
],
"score": 0.92,
"content": "y ^ { ( j ) } \\in \\{ 0 , 1 , . . . , N _ { c l } - 1 \\}",
"type": "inline_equation"
},
{
"bbox": [
429,
261,
506,
275
],
"score": 1.0,
"content": "is the true label of",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
107,
273,
409,
288
],
"spans": [
{
"bbox": [
107,
275,
122,
286
],
"score": 0.87,
"content": "N _ { c l }",
"type": "inline_equation"
},
{
"bbox": [
123,
273,
180,
288
],
"score": 1.0,
"content": "MI tasks, and",
"type": "text"
},
{
"bbox": [
180,
274,
250,
285
],
"score": 0.92,
"content": "\\mathbf { X } ^ { ( j ) } \\in \\mathbb { R } ^ { N _ { s } \\times N _ { c h } }",
"type": "inline_equation"
},
{
"bbox": [
250,
273,
409,
288
],
"score": 1.0,
"content": "the multi-channel recording defined as",
"type": "text"
}
],
"index": 13
}
],
"index": 11.5,
"bbox_fs": [
105,
238,
506,
288
]
},
{
"type": "interline_equation",
"bbox": [
236,
287,
374,
308
],
"lines": [
{
"bbox": [
236,
287,
374,
308
],
"spans": [
{
"bbox": [
236,
287,
374,
308
],
"score": 0.94,
"content": "\\mathbf { X } ^ { ( j ) } : = \\left( \\mathbf { x } _ { 0 } ^ { ( j ) } , \\mathbf { x } _ { 1 } ^ { ( j ) } , . . . , \\mathbf { x } _ { N _ { c h } - 1 } ^ { ( j ) } \\right) ,",
"type": "interline_equation",
"image_path": "4adef2bd32024f0155b08bd021946820057f4e0a6d08b9b081db9e63f5cb922a.jpg"
}
]
}
],
"index": 14,
"virtual_lines": [
{
"bbox": [
236,
287,
374,
308
],
"spans": [],
"index": 14
}
]
},
{
"type": "text",
"bbox": [
109,
312,
505,
336
],
"lines": [
{
"bbox": [
107,
309,
504,
325
],
"spans": [
{
"bbox": [
107,
309,
126,
325
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
127,
309,
173,
325
],
"score": 0.93,
"content": "\\mathbf { x } _ { i } ^ { ( j ) } \\in \\mathbb { R } ^ { N _ { s } }",
"type": "inline_equation"
},
{
"bbox": [
174,
309,
326,
325
],
"score": 1.0,
"content": "corresponding to the recording of the",
"type": "text"
},
{
"bbox": [
327,
313,
332,
324
],
"score": 0.82,
"content": "j",
"type": "inline_equation"
},
{
"bbox": [
333,
309,
394,
325
],
"score": 1.0,
"content": "-th trial and the",
"type": "text"
},
{
"bbox": [
395,
313,
399,
322
],
"score": 0.78,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
400,
309,
490,
325
],
"score": 1.0,
"content": "-th channel containing",
"type": "text"
},
{
"bbox": [
491,
313,
504,
323
],
"score": 0.85,
"content": "N _ { s }",
"type": "inline_equation"
}
],
"index": 15
},
{
"bbox": [
107,
323,
390,
337
],
"spans": [
{
"bbox": [
107,
323,
285,
337
],
"score": 1.0,
"content": "temporal samples. For simplicity, we denote",
"type": "text"
},
{
"bbox": [
286,
324,
330,
335
],
"score": 0.91,
"content": "\\mathbf { X } : = \\mathbf { X } ^ { ( j ) }",
"type": "inline_equation"
},
{
"bbox": [
330,
323,
349,
337
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
349,
324,
387,
336
],
"score": 0.93,
"content": "y : = y ^ { ( j ) }",
"type": "inline_equation"
},
{
"bbox": [
387,
323,
390,
337
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 16
}
],
"index": 15.5,
"bbox_fs": [
107,
309,
504,
337
]
},
{
"type": "text",
"bbox": [
106,
341,
503,
364
],
"lines": [
{
"bbox": [
106,
341,
505,
354
],
"spans": [
{
"bbox": [
106,
341,
505,
354
],
"score": 1.0,
"content": "The EEG recordings are often preprocessed with a band-pass filter, e.g., using a Fast Fourier Transform",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
351,
344,
366
],
"spans": [
{
"bbox": [
105,
351,
154,
366
],
"score": 1.0,
"content": "(FFT) filter",
"type": "text"
},
{
"bbox": [
154,
353,
179,
365
],
"score": 0.92,
"content": "h _ { b p } ( \\cdot )",
"type": "inline_equation"
},
{
"bbox": [
179,
351,
344,
366
],
"score": 1.0,
"content": ", before being fed to a classifier, yielding",
"type": "text"
}
],
"index": 18
}
],
"index": 17.5,
"bbox_fs": [
105,
341,
505,
366
]
},
{
"type": "interline_equation",
"bbox": [
191,
365,
420,
379
],
"lines": [
{
"bbox": [
191,
365,
420,
379
],
"spans": [
{
"bbox": [
191,
365,
420,
379
],
"score": 0.88,
"content": "\\mathbf { X } _ { b p } = H _ { b p } ( \\mathbf { X } ) = \\left( h _ { b p } ( \\mathbf { x } _ { 0 } ) , h _ { b p } ( \\mathbf { x } _ { 1 } ) , . . . , h _ { b p } ( \\mathbf { x } _ { N _ { c h } - 1 } ) \\right) .",
"type": "interline_equation",
"image_path": "92578742fa44b8eb625a3c6e4e087ff7ad891f1315813fccd22129d3165cbf68.jpg"
}
]
}
],
"index": 19,
"virtual_lines": [
{
"bbox": [
191,
365,
420,
379
],
"spans": [],
"index": 19
}
]
},
{
"type": "text",
"bbox": [
107,
380,
506,
414
],
"lines": [
{
"bbox": [
105,
378,
505,
393
],
"spans": [
{
"bbox": [
105,
378,
237,
393
],
"score": 1.0,
"content": "Finally, the preprocessed signal",
"type": "text"
},
{
"bbox": [
238,
380,
256,
392
],
"score": 0.9,
"content": "\\mathbf { X } _ { b p }",
"type": "inline_equation"
},
{
"bbox": [
256,
378,
401,
393
],
"score": 1.0,
"content": "is classified with a trainable model",
"type": "text"
},
{
"bbox": [
402,
380,
409,
391
],
"score": 0.84,
"content": "f",
"type": "inline_equation"
},
{
"bbox": [
409,
378,
482,
393
],
"score": 1.0,
"content": "and is mapped to",
"type": "text"
},
{
"bbox": [
482,
381,
505,
392
],
"score": 0.82,
"content": "\\mathbf { p } : =",
"type": "inline_equation"
}
],
"index": 20
},
{
"bbox": [
107,
390,
506,
405
],
"spans": [
{
"bbox": [
107,
391,
142,
404
],
"score": 0.89,
"content": "f \\left( { \\bf { X } } _ { b p } \\right)",
"type": "inline_equation"
},
{
"bbox": [
142,
390,
171,
405
],
"score": 1.0,
"content": ", where",
"type": "text"
},
{
"bbox": [
171,
391,
210,
403
],
"score": 0.91,
"content": "\\mathbf { p } \\in \\mathbb { R } ^ { N _ { c l } }",
"type": "inline_equation"
},
{
"bbox": [
210,
390,
506,
405
],
"score": 1.0,
"content": "contains the output probabilities, e.g., originating from a softmax activation",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
402,
494,
416
],
"spans": [
{
"bbox": [
105,
402,
187,
416
],
"score": 1.0,
"content": "as final operation in",
"type": "text"
},
{
"bbox": [
188,
403,
195,
414
],
"score": 0.84,
"content": "f",
"type": "inline_equation"
},
{
"bbox": [
195,
402,
314,
416
],
"score": 1.0,
"content": ". The model’s final prediction",
"type": "text"
},
{
"bbox": [
315,
403,
321,
414
],
"score": 0.85,
"content": "\\hat { y }",
"type": "inline_equation"
},
{
"bbox": [
321,
402,
482,
416
],
"score": 1.0,
"content": "is the index with the maximum score in",
"type": "text"
},
{
"bbox": [
482,
404,
489,
414
],
"score": 0.33,
"content": "\\mathbf { p }",
"type": "inline_equation"
},
{
"bbox": [
490,
402,
494,
416
],
"score": 1.0,
"content": ":",
"type": "text"
}
],
"index": 22
}
],
"index": 21,
"bbox_fs": [
105,
378,
506,
416
]
},
{
"type": "interline_equation",
"bbox": [
219,
415,
391,
439
],
"lines": [
{
"bbox": [
219,
415,
391,
439
],
"spans": [
{
"bbox": [
219,
415,
391,
439
],
"score": 0.93,
"content": "\\boldsymbol { \\hat { y } } = \\boldsymbol { \\hat { f } } \\left( \\mathbf { X } _ { b p } \\right) = \\underset { y \\in \\{ 0 , \\ldots , N _ { c l } - 1 \\} } { \\mathrm { a r g m a x } } ~ f \\left( \\mathbf { X } _ { b p } \\right) [ y ] .",
"type": "interline_equation",
"image_path": "090de2464ceebb19449b8aaa8df74dfb9fcbf1932b76ec9c09e0616b0fd63a0c.jpg"
}
]
}
],
"index": 23,
"virtual_lines": [
{
"bbox": [
219,
415,
391,
439
],
"spans": [],
"index": 23
}
]
},
{
"type": "title",
"bbox": [
108,
450,
248,
461
],
"lines": [
{
"bbox": [
106,
450,
249,
462
],
"spans": [
{
"bbox": [
106,
450,
249,
462
],
"score": 1.0,
"content": "2.2 INSTANCE-BASED ATTACKS",
"type": "text"
}
],
"index": 24
}
],
"index": 24
},
{
"type": "text",
"bbox": [
106,
470,
505,
516
],
"lines": [
{
"bbox": [
106,
470,
505,
483
],
"spans": [
{
"bbox": [
106,
470,
315,
483
],
"score": 1.0,
"content": "Instance-based attacks try to fool an EEG classifier",
"type": "text"
},
{
"bbox": [
315,
471,
322,
482
],
"score": 0.86,
"content": "f",
"type": "inline_equation"
},
{
"bbox": [
322,
470,
442,
483
],
"score": 1.0,
"content": "to misclassify an EEG signal",
"type": "text"
},
{
"bbox": [
442,
471,
452,
481
],
"score": 0.58,
"content": "\\mathbf { X }",
"type": "inline_equation"
},
{
"bbox": [
452,
470,
505,
483
],
"score": 1.0,
"content": "to a targeted",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
481,
505,
493
],
"spans": [
{
"bbox": [
106,
481,
128,
493
],
"score": 1.0,
"content": "class",
"type": "text"
},
{
"bbox": [
129,
483,
138,
493
],
"score": 0.83,
"content": "y _ { t }",
"type": "inline_equation"
},
{
"bbox": [
138,
481,
398,
493
],
"score": 1.0,
"content": ". In this section, we describe the attack directly on the classifier",
"type": "text"
},
{
"bbox": [
398,
482,
405,
493
],
"score": 0.85,
"content": "f",
"type": "inline_equation"
},
{
"bbox": [
406,
481,
505,
493
],
"score": 1.0,
"content": "without considering the",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
493,
506,
505
],
"spans": [
{
"bbox": [
105,
493,
166,
505
],
"score": 1.0,
"content": "preprocessing",
"type": "text"
},
{
"bbox": [
166,
493,
183,
505
],
"score": 0.91,
"content": "H _ { b p }",
"type": "inline_equation"
},
{
"bbox": [
184,
493,
506,
505
],
"score": 1.0,
"content": "; the inclusion of the preprocessing is described in Section 3.3. We define an",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
104,
503,
365,
517
],
"spans": [
{
"bbox": [
104,
503,
216,
517
],
"score": 1.0,
"content": "adversarial example as any",
"type": "text"
},
{
"bbox": [
216,
504,
324,
515
],
"score": 0.89,
"content": "\\mathbf { X } ^ { * } = \\mathbf { X } + \\mathbf { V } \\in \\mathbb { R } ^ { N _ { s } \\times N _ { c h } }",
"type": "inline_equation"
},
{
"bbox": [
324,
503,
365,
517
],
"score": 1.0,
"content": "such that",
"type": "text"
}
],
"index": 28
}
],
"index": 26.5,
"bbox_fs": [
104,
470,
506,
517
]
},
{
"type": "interline_equation",
"bbox": [
259,
517,
351,
533
],
"lines": [
{
"bbox": [
259,
517,
351,
533
],
"spans": [
{
"bbox": [
259,
517,
351,
533
],
"score": 0.88,
"content": "\\hat { f } \\left( \\mathbf { X } \\right) \\neq \\hat { f } \\left( \\mathbf { X } ^ { * } \\right) = y _ { t } .",
"type": "interline_equation",
"image_path": "ae96fbca25561b728f407743e94ef7daae4c068465648eca1989b79f54f18f8f.jpg"
}
]
}
],
"index": 29,
"virtual_lines": [
{
"bbox": [
259,
517,
351,
533
],
"spans": [],
"index": 29
}
]
},
{
"type": "text",
"bbox": [
106,
541,
503,
563
],
"lines": [
{
"bbox": [
104,
539,
503,
554
],
"spans": [
{
"bbox": [
104,
539,
443,
554
],
"score": 1.0,
"content": "FGSM. The FGSM (Goodfellow et al., 2015) generates an adversarial perturbation",
"type": "text"
},
{
"bbox": [
443,
540,
503,
551
],
"score": 0.9,
"content": "\\mathbf { V } \\in \\mathbb { R } ^ { N _ { s } \\times N _ { c h } }",
"type": "inline_equation"
}
],
"index": 30
},
{
"bbox": [
105,
551,
437,
565
],
"spans": [
{
"bbox": [
105,
551,
162,
565
],
"score": 1.0,
"content": "of magnitude",
"type": "text"
},
{
"bbox": [
162,
554,
167,
562
],
"score": 0.67,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
168,
551,
437,
565
],
"score": 1.0,
"content": "which points in the negative direction of a loss function’s gradient:",
"type": "text"
}
],
"index": 31
}
],
"index": 30.5,
"bbox_fs": [
104,
539,
503,
565
]
},
{
"type": "interline_equation",
"bbox": [
239,
564,
370,
578
],
"lines": [
{
"bbox": [
239,
564,
370,
578
],
"spans": [
{
"bbox": [
239,
564,
370,
578
],
"score": 0.9,
"content": "\\mathbf { V } = - \\epsilon \\cdot \\mathrm { s i g n } \\left( \\nabla _ { \\mathbf { X } } \\cdot l \\left( \\mathbf { X } , y _ { t } \\right) \\right) ,",
"type": "interline_equation",
"image_path": "0d4290c3afb9da2ee94681d4b078c69233401f350f53f236fd055d34090accc5.jpg"
}
]
}
],
"index": 32,
"virtual_lines": [
{
"bbox": [
239,
564,
370,
578
],
"spans": [],
"index": 32
}
]
},
{
"type": "text",
"bbox": [
106,
578,
347,
590
],
"lines": [
{
"bbox": [
105,
577,
345,
592
],
"spans": [
{
"bbox": [
105,
577,
345,
592
],
"score": 1.0,
"content": "where the loss function contains the negative log likelihood",
"type": "text"
}
],
"index": 33
}
],
"index": 33,
"bbox_fs": [
105,
577,
345,
592
]
},
{
"type": "interline_equation",
"bbox": [
212,
591,
398,
604
],
"lines": [
{
"bbox": [
212,
591,
398,
604
],
"spans": [
{
"bbox": [
212,
591,
398,
604
],
"score": 0.9,
"content": "l ( \\mathbf { X } , y _ { t } ) = - \\log \\left( { \\mathbf { p } } [ y _ { t } ] \\right) = - \\log \\left( f \\left( \\mathbf { X } \\right) [ y _ { t } ] \\right) .",
"type": "interline_equation",
"image_path": "76d960042d0c715217ff55097d0087d7852aed951636159758fb7c8f7d3fa671.jpg"
}
]
}
],
"index": 34,
"virtual_lines": [
{
"bbox": [
212,
591,
398,
604
],
"spans": [],
"index": 34
}
]
},
{
"type": "text",
"bbox": [
105,
606,
504,
628
],
"lines": [
{
"bbox": [
105,
605,
505,
619
],
"spans": [
{
"bbox": [
105,
605,
119,
619
],
"score": 1.0,
"content": "As",
"type": "text"
},
{
"bbox": [
119,
607,
127,
617
],
"score": 0.37,
"content": "\\mathbf { p }",
"type": "inline_equation"
},
{
"bbox": [
128,
605,
505,
619
],
"score": 1.0,
"content": "is the output of the softmax activation function, equation 6 becomes a cross-entropy loss which",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
616,
382,
629
],
"spans": [
{
"bbox": [
105,
616,
194,
629
],
"score": 1.0,
"content": "maximizes the output",
"type": "text"
},
{
"bbox": [
195,
616,
216,
628
],
"score": 0.91,
"content": "\\mathbf { p } [ y _ { t } ]",
"type": "inline_equation"
},
{
"bbox": [
217,
616,
382,
629
],
"score": 1.0,
"content": "while minimizing the remaining outputs.",
"type": "text"
}
],
"index": 36
}
],
"index": 35.5,
"bbox_fs": [
105,
605,
505,
629
]
},
{
"type": "text",
"bbox": [
106,
638,
505,
694
],
"lines": [
{
"bbox": [
105,
638,
506,
652
],
"spans": [
{
"bbox": [
105,
638,
506,
652
],
"score": 1.0,
"content": "PGD. The projected gradient descent (PGD) (Madry et al., 2018) is a variant of the basic iterative",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
650,
506,
662
],
"spans": [
{
"bbox": [
106,
650,
506,
662
],
"score": 1.0,
"content": "method (Kurakin et al.), generally considered to be more effective than FGSM. PGD aims to find a",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
660,
506,
675
],
"spans": [
{
"bbox": [
105,
660,
316,
675
],
"score": 1.0,
"content": "perturbation by iteratively taking small steps of size",
"type": "text"
},
{
"bbox": [
316,
663,
324,
671
],
"score": 0.79,
"content": "\\alpha",
"type": "inline_equation"
},
{
"bbox": [
325,
660,
506,
675
],
"score": 1.0,
"content": "in the gradient’s direction and projecting the",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
672,
505,
684
],
"spans": [
{
"bbox": [
106,
672,
505,
684
],
"score": 1.0,
"content": "resulting perturbation back to the sample’s neighborhood after each iteration. We randomly initialize",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
682,
472,
696
],
"spans": [
{
"bbox": [
106,
682,
188,
696
],
"score": 1.0,
"content": "the attack inside the",
"type": "text"
},
{
"bbox": [
189,
684,
205,
694
],
"score": 0.9,
"content": "L _ { \\infty }",
"type": "inline_equation"
},
{
"bbox": [
205,
682,
261,
696
],
"score": 1.0,
"content": "ball of radius",
"type": "text"
},
{
"bbox": [
261,
685,
267,
693
],
"score": 0.73,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
267,
682,
354,
696
],
"score": 1.0,
"content": "and update the attack",
"type": "text"
},
{
"bbox": [
355,
683,
378,
695
],
"score": 0.91,
"content": "\\mathbf { V } _ { t + 1 }",
"type": "inline_equation"
},
{
"bbox": [
378,
682,
446,
696
],
"score": 1.0,
"content": "for any iteration",
"type": "text"
},
{
"bbox": [
446,
684,
451,
693
],
"score": 0.76,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
451,
682,
472,
696
],
"score": 1.0,
"content": "with",
"type": "text"
}
],
"index": 41
}
],
"index": 39,
"bbox_fs": [
105,
638,
506,
696
]
},
{
"type": "interline_equation",
"bbox": [
200,
696,
408,
709
],
"lines": [
{
"bbox": [
200,
696,
408,
709
],
"spans": [
{
"bbox": [
200,
696,
408,
709
],
"score": 0.83,
"content": "\\mathbf { V } _ { t + 1 } = \\mathrm { c l i p } _ { \\epsilon } \\left( \\mathbf { V } _ { t } - \\alpha \\cdot \\mathrm { s i g n } \\left( \\nabla _ { \\mathbf { V } } l \\left( \\mathbf { X } + \\mathbf { V } _ { t } , y _ { t } \\right) \\right) \\right) ,",
"type": "interline_equation",
"image_path": "5018cbeb47f39276c0e84f65e695e567cca2809727b109293d37f372e63a064d.jpg"
}
]
}
],
"index": 42,
"virtual_lines": [
{
"bbox": [
200,
696,
408,
709
],
"spans": [],
"index": 42
}
]
},
{
"type": "text",
"bbox": [
106,
710,
507,
732
],
"lines": [
{
"bbox": [
106,
709,
504,
721
],
"spans": [
{
"bbox": [
106,
709,
134,
721
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
712,
142,
720
],
"score": 0.78,
"content": "\\alpha",
"type": "inline_equation"
},
{
"bbox": [
142,
709,
254,
721
],
"score": 1.0,
"content": "is a step size smaller than",
"type": "text"
},
{
"bbox": [
254,
712,
260,
720
],
"score": 0.68,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
260,
709,
504,
721
],
"score": 1.0,
"content": "which decays linearly with each iteration and the function",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
720,
353,
734
],
"spans": [
{
"bbox": [
106,
720,
139,
733
],
"score": 0.87,
"content": "\\mathrm { c l i p } _ { \\epsilon } \\left( \\cdot \\right)",
"type": "inline_equation"
},
{
"bbox": [
139,
720,
343,
734
],
"score": 1.0,
"content": "clips the signal at the maximum desired amplitude",
"type": "text"
},
{
"bbox": [
344,
723,
349,
730
],
"score": 0.72,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
349,
720,
353,
734
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 44
}
],
"index": 43.5,
"bbox_fs": [
106,
709,
504,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
108,
82,
223,
93
],
"lines": [
{
"bbox": [
106,
82,
225,
95
],
"spans": [
{
"bbox": [
106,
82,
225,
95
],
"score": 1.0,
"content": "2.3 UNIVERSAL ATTACKS",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
102,
506,
136
],
"lines": [
{
"bbox": [
105,
101,
506,
117
],
"spans": [
{
"bbox": [
105,
101,
506,
117
],
"score": 1.0,
"content": "UAPs have been introduced by Moosavi-Dezfooli et al. (2017) in the context of natural images,",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
114,
506,
126
],
"spans": [
{
"bbox": [
105,
114,
506,
126
],
"score": 1.0,
"content": "seeking to find an image-agnostic perturbation that fools the classifier on any input image. In the BCI",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
123,
433,
138
],
"spans": [
{
"bbox": [
105,
123,
331,
138
],
"score": 1.0,
"content": "domain (Liu et al., 2021), we seek to find a perturbation",
"type": "text"
},
{
"bbox": [
332,
124,
391,
135
],
"score": 0.91,
"content": "\\mathbf { V } \\in \\mathbb { R } ^ { N _ { s } \\times N _ { c h } }",
"type": "inline_equation"
},
{
"bbox": [
392,
123,
433,
138
],
"score": 1.0,
"content": "such that",
"type": "text"
}
],
"index": 3
}
],
"index": 2
},
{
"type": "interline_equation",
"bbox": [
222,
140,
388,
155
],
"lines": [
{
"bbox": [
222,
140,
388,
155
],
"spans": [
{
"bbox": [
222,
140,
388,
155
],
"score": 0.89,
"content": "\\begin{array} { r } { \\hat { f } \\left( \\mathbf { X } + \\mathbf { V } \\right) \\neq \\hat { f } \\left( \\mathbf { X } \\right) \\mathrm { f o r } ^ { * } \\mathrm { m o s t } ^ { * } \\mathbf { X } \\sim D , } \\end{array}",
"type": "interline_equation",
"image_path": "831387fcf3fae8922d58bc63c66078fe603ae4b2c440daa59c8a14c9c37671c9.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
222,
140,
388,
155
],
"spans": [],
"index": 4
}
]
},
{
"type": "text",
"bbox": [
106,
158,
507,
181
],
"lines": [
{
"bbox": [
106,
158,
505,
171
],
"spans": [
{
"bbox": [
106,
158,
132,
171
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
133,
159,
142,
168
],
"score": 0.83,
"content": "D",
"type": "inline_equation"
},
{
"bbox": [
143,
158,
505,
171
],
"score": 1.0,
"content": "is the distribution of the EEG data. The UAP can be determined by optimizing the negative",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
169,
495,
182
],
"spans": [
{
"bbox": [
106,
169,
244,
182
],
"score": 1.0,
"content": "log-likelihood loss with respect to",
"type": "text"
},
{
"bbox": [
245,
169,
255,
180
],
"score": 0.69,
"content": "\\mathbf { V }",
"type": "inline_equation"
},
{
"bbox": [
255,
169,
495,
182
],
"score": 1.0,
"content": "using batch gradient descent on the trials in the training set.",
"type": "text"
}
],
"index": 6
}
],
"index": 5.5
},
{
"type": "title",
"bbox": [
107,
196,
333,
209
],
"lines": [
{
"bbox": [
105,
195,
334,
210
],
"spans": [
{
"bbox": [
105,
195,
334,
210
],
"score": 1.0,
"content": "3 MODELING PRACTICAL ATTACKS IN BCI",
"type": "text"
}
],
"index": 7
}
],
"index": 7
},
{
"type": "text",
"bbox": [
106,
220,
506,
342
],
"lines": [
{
"bbox": [
105,
220,
507,
234
],
"spans": [
{
"bbox": [
105,
220,
507,
234
],
"score": 1.0,
"content": "This section is the main contribution of the paper: we present a design of practical DoS attacks on MI-",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
231,
505,
245
],
"spans": [
{
"bbox": [
105,
231,
505,
245
],
"score": 1.0,
"content": "BCIs that operates at the source of the signal acquisition. We propose a new method to eliminate the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
243,
506,
255
],
"spans": [
{
"bbox": [
105,
243,
506,
255
],
"score": 1.0,
"content": "square wave artifacts to generate adversarial examples that are natural and physiologically plausible.",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
253,
505,
267
],
"spans": [
{
"bbox": [
105,
253,
505,
267
],
"score": 1.0,
"content": "The perturbation is emitted by a smart, adversarial device placed close to the ear, e.g., a smart glass or",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
265,
505,
278
],
"spans": [
{
"bbox": [
105,
265,
505,
278
],
"score": 1.0,
"content": "in-ear headphones, and is propagated to the individual EEG electrodes over the scalp’s skin. As can",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
276,
506,
288
],
"spans": [
{
"bbox": [
105,
276,
506,
288
],
"score": 1.0,
"content": "be experimentally observed on measured EEG traces (Merlet et al., 2013; Sazgar & Young, 2019),",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
287,
506,
300
],
"spans": [
{
"bbox": [
106,
287,
506,
300
],
"score": 1.0,
"content": "the same electrical source, e.g., electrocardiographic activities, is sensed by each EEG electrode with",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
298,
506,
311
],
"spans": [
{
"bbox": [
106,
298,
506,
311
],
"score": 1.0,
"content": "different degrees of attenuation and delay. We present a practical propagation model that determines",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
309,
505,
320
],
"spans": [
{
"bbox": [
106,
309,
505,
320
],
"score": 1.0,
"content": "the magnitude and delay for every individual electrode based on the distance along the scalp to the",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
320,
506,
333
],
"spans": [
{
"bbox": [
106,
320,
506,
333
],
"score": 1.0,
"content": "adversarial device. The perturbation is trained end-to-end to fool the classifier to always output “rest,”",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
330,
506,
343
],
"spans": [
{
"bbox": [
105,
330,
506,
343
],
"score": 1.0,
"content": "hence DoS, while respecting the spatial model and the amplitude constraints to remain imperceptible.",
"type": "text"
}
],
"index": 18
}
],
"index": 13
},
{
"type": "title",
"bbox": [
107,
354,
437,
367
],
"lines": [
{
"bbox": [
105,
354,
438,
368
],
"spans": [
{
"bbox": [
105,
354,
438,
368
],
"score": 1.0,
"content": "3.1 DESIGN AND ASSESSMENT OF PHYSIOLOGICALLY PLAUSIBLE ATTACKS",
"type": "text"
}
],
"index": 19
}
],
"index": 19
},
{
"type": "text",
"bbox": [
106,
375,
506,
454
],
"lines": [
{
"bbox": [
105,
375,
506,
388
],
"spans": [
{
"bbox": [
105,
375,
506,
388
],
"score": 1.0,
"content": "PGD-designed attacks on EEG tend to form perturbation signals which resemble a square-wave",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
104,
386,
507,
400
],
"spans": [
{
"bbox": [
104,
386,
507,
400
],
"score": 1.0,
"content": "artifact (see Figure 2), an effect that has been observed on ECG data, too (Han et al., 2020). However,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
397,
506,
410
],
"spans": [
{
"bbox": [
105,
397,
506,
410
],
"score": 1.0,
"content": "EEG signals are of random nature and can be modeled as frequency dependent stationary or non-",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
408,
506,
421
],
"spans": [
{
"bbox": [
105,
408,
506,
421
],
"score": 1.0,
"content": "stationary random processes (Karlekar & Gupta, 2014). To this end, we introduce a new loss term in",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
420,
505,
432
],
"spans": [
{
"bbox": [
106,
420,
505,
432
],
"score": 1.0,
"content": "the PGD optimization such that the perturbation resembles the random nature of EEG signals, which",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
430,
505,
443
],
"spans": [
{
"bbox": [
105,
430,
505,
443
],
"score": 1.0,
"content": "we achieve by promoting signal changes represented in the first order derivative. We estimate the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
103,
439,
505,
456
],
"spans": [
{
"bbox": [
103,
439,
197,
456
],
"score": 1.0,
"content": "per-channel derivative",
"type": "text"
},
{
"bbox": [
197,
441,
365,
455
],
"score": 0.92,
"content": "\\begin{array} { r } { \\bar { { \\bf V } ^ { \\prime } } = ( { \\bf v } _ { 0 } ^ { \\prime } , { \\bf v } _ { 1 } ^ { \\prime } , . . . , { \\bf v } _ { N _ { c h } - 1 } ^ { \\prime } ) \\in \\mathbb { R } ^ { N _ { s } - 1 \\times N _ { c h } } } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
365,
439,
505,
456
],
"score": 1.0,
"content": "using the sample-wise difference:",
"type": "text"
}
],
"index": 26
}
],
"index": 23
},
{
"type": "interline_equation",
"bbox": [
155,
458,
456,
471
],
"lines": [
{
"bbox": [
155,
458,
456,
471
],
"spans": [
{
"bbox": [
155,
458,
456,
471
],
"score": 0.83,
"content": "\\begin{array} { r } { \\mathbf { v } _ { c } ^ { \\prime } [ t ] : = \\mathbf { v } _ { c } [ t ] - \\mathbf { v } _ { c } [ t - 1 ] \\quad t \\in \\{ 1 , 2 , . . . , N _ { s } - 1 \\} , c \\in \\{ 0 , 1 , . . . , N _ { c h } - 1 \\} } \\end{array}",
"type": "interline_equation",
"image_path": "ce0c71c6986adead5a7ef4f07c8039c9cf9397e84e03cd7c9175ee0546e8ce8a.jpg"
}
]
}
],
"index": 27,
"virtual_lines": [
{
"bbox": [
155,
458,
456,
471
],
"spans": [],
"index": 27
}
]
},
{
"type": "text",
"bbox": [
106,
476,
505,
515
],
"lines": [
{
"bbox": [
105,
472,
508,
496
],
"spans": [
{
"bbox": [
105,
477,
262,
492
],
"score": 1.0,
"content": "The additive loss term is determined by",
"type": "text"
},
{
"bbox": [
262,
475,
370,
491
],
"score": 0.91,
"content": "\\begin{array} { r } { l _ { 1 } ( { \\bf V } ) = - \\frac { \\beta } { \\epsilon } \\sum _ { c = 1 } ^ { N _ { c h } } | | { \\bf v } _ { c } ^ { \\prime } | | _ { 1 } } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
303,
472,
508,
496
],
"score": 1.0,
"content": "− β\u000f PNchc=1 ||v0c||1, where || · ||1 is the `1-norm, \u000f the",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
489,
505,
502
],
"spans": [
{
"bbox": [
106,
489,
260,
502
],
"score": 1.0,
"content": "maximum perturbation amplitude, and",
"type": "text"
},
{
"bbox": [
261,
490,
286,
500
],
"score": 0.9,
"content": "\\beta \\geq 0",
"type": "inline_equation"
},
{
"bbox": [
286,
489,
505,
502
],
"score": 1.0,
"content": "a weighting factor. When designing a one-dimensional",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
500,
347,
515
],
"spans": [
{
"bbox": [
105,
500,
271,
515
],
"score": 1.0,
"content": "perturbation, the derivative loss becomes",
"type": "text"
},
{
"bbox": [
272,
500,
343,
515
],
"score": 0.94,
"content": "\\begin{array} { r } { l _ { 1 } ( \\mathbf { v } ) = - \\frac { \\beta } { \\epsilon } | | \\mathbf { v } ^ { \\prime } | | } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
343,
500,
347,
515
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 30
}
],
"index": 29
},
{
"type": "text",
"bbox": [
107,
524,
505,
580
],
"lines": [
{
"bbox": [
105,
524,
505,
538
],
"spans": [
{
"bbox": [
105,
524,
505,
538
],
"score": 1.0,
"content": "Measuring the Plausibility of Attacks None of the previous works have given quantitative",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
536,
505,
548
],
"spans": [
{
"bbox": [
105,
536,
505,
548
],
"score": 1.0,
"content": "measures to assess the physiological plausibility of an EEG adversarial attack. In this work, we",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
547,
505,
559
],
"spans": [
{
"bbox": [
105,
547,
505,
559
],
"score": 1.0,
"content": "propose data-driven measures for quantifying the naturalism of an attack. We compute either the",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
558,
504,
570
],
"spans": [
{
"bbox": [
106,
558,
504,
570
],
"score": 1.0,
"content": "cross correlation, the Euclidian distance, or the cosine similarity between the attacked signal and the",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
569,
491,
581
],
"spans": [
{
"bbox": [
106,
569,
291,
581
],
"score": 1.0,
"content": "original EEG, and average the values over the",
"type": "text"
},
{
"bbox": [
291,
569,
309,
580
],
"score": 0.91,
"content": "N _ { c h }",
"type": "inline_equation"
},
{
"bbox": [
309,
569,
491,
581
],
"score": 1.0,
"content": "channels and over the samples in the dataset.",
"type": "text"
}
],
"index": 35
}
],
"index": 33
},
{
"type": "title",
"bbox": [
108,
593,
262,
604
],
"lines": [
{
"bbox": [
106,
593,
264,
606
],
"spans": [
{
"bbox": [
106,
593,
264,
606
],
"score": 1.0,
"content": "3.2 SPATIAL PROPAGATION MODEL",
"type": "text"
}
],
"index": 36
}
],
"index": 36
},
{
"type": "text",
"bbox": [
106,
613,
505,
713
],
"lines": [
{
"bbox": [
106,
613,
506,
627
],
"spans": [
{
"bbox": [
106,
613,
506,
627
],
"score": 1.0,
"content": "So far, a perturbation signal was designed for every individual channel. It is unrealistic for an attacker",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
624,
505,
637
],
"spans": [
{
"bbox": [
105,
624,
505,
637
],
"score": 1.0,
"content": "to perturb the signal for all individual channels simultaneously; hence, we consider a more practical",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
104,
633,
506,
649
],
"spans": [
{
"bbox": [
104,
633,
256,
649
],
"score": 1.0,
"content": "use case where the perturbation signal",
"type": "text"
},
{
"bbox": [
257,
635,
293,
646
],
"score": 0.91,
"content": "\\mathbf { v } \\in \\mathbb { R } ^ { N _ { s } }",
"type": "inline_equation"
},
{
"bbox": [
293,
633,
506,
649
],
"score": 1.0,
"content": "is emitted from one location, e.g., from an adversarial",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
645,
505,
659
],
"spans": [
{
"bbox": [
105,
645,
505,
659
],
"score": 1.0,
"content": "device placed on the left side of the subject or close to the left ear. More specifically, in this study, we",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
658,
506,
671
],
"spans": [
{
"bbox": [
105,
658,
506,
671
],
"score": 1.0,
"content": "assume that the EEG electrode at the position T9 according to the international 10-10 system (Sch),",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
667,
505,
683
],
"spans": [
{
"bbox": [
105,
667,
505,
683
],
"score": 1.0,
"content": "which is the closest to the left ear, senses the largest perturbation. The signal subsequently propagates",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
678,
505,
693
],
"spans": [
{
"bbox": [
105,
678,
505,
693
],
"score": 1.0,
"content": "over the skin to each electrode, which results in an individual magnitude and delay depending on",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
690,
506,
703
],
"spans": [
{
"bbox": [
106,
690,
506,
703
],
"score": 1.0,
"content": "the distance between the adversarial device and the electrode. More formally, we model the sensed",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
702,
290,
713
],
"spans": [
{
"bbox": [
105,
702,
201,
713
],
"score": 1.0,
"content": "perturbation at channel",
"type": "text"
},
{
"bbox": [
201,
702,
205,
711
],
"score": 0.75,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
206,
702,
273,
713
],
"score": 1.0,
"content": "and time instant",
"type": "text"
},
{
"bbox": [
273,
703,
278,
711
],
"score": 0.8,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
278,
702,
290,
713
],
"score": 1.0,
"content": "as",
"type": "text"
}
],
"index": 45
}
],
"index": 41
},
{
"type": "interline_equation",
"bbox": [
204,
716,
406,
730
],
"lines": [
{
"bbox": [
204,
716,
406,
730
],
"spans": [
{
"bbox": [
204,
716,
406,
730
],
"score": 0.88,
"content": "h _ { i } ( \\mathbf { v } , \\lambda _ { m } , \\lambda _ { d } ) ( t ) : = m ( l _ { i } , \\lambda _ { m } ) \\cdot \\mathbf { v } \\left( t - d ( l _ { i } , \\lambda _ { d } ) \\right) ,",
"type": "interline_equation",
"image_path": "376c0312852c9f41d7a3778ae583d8e40b0cf5932a3541026cd26602e02373b7.jpg"
}
]
}
],
"index": 46,
"virtual_lines": [
{
"bbox": [
204,
716,
406,
730
],
"spans": [],
"index": 46
}
]
}
],
"page_idx": 3,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
26,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
309,
38
],
"spans": [
{
"bbox": [
106,
25,
309,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
752,
308,
759
],
"lines": []
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
108,
82,
223,
93
],
"lines": [
{
"bbox": [
106,
82,
225,
95
],
"spans": [
{
"bbox": [
106,
82,
225,
95
],
"score": 1.0,
"content": "2.3 UNIVERSAL ATTACKS",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
106,
102,
506,
136
],
"lines": [
{
"bbox": [
105,
101,
506,
117
],
"spans": [
{
"bbox": [
105,
101,
506,
117
],
"score": 1.0,
"content": "UAPs have been introduced by Moosavi-Dezfooli et al. (2017) in the context of natural images,",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
114,
506,
126
],
"spans": [
{
"bbox": [
105,
114,
506,
126
],
"score": 1.0,
"content": "seeking to find an image-agnostic perturbation that fools the classifier on any input image. In the BCI",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
123,
433,
138
],
"spans": [
{
"bbox": [
105,
123,
331,
138
],
"score": 1.0,
"content": "domain (Liu et al., 2021), we seek to find a perturbation",
"type": "text"
},
{
"bbox": [
332,
124,
391,
135
],
"score": 0.91,
"content": "\\mathbf { V } \\in \\mathbb { R } ^ { N _ { s } \\times N _ { c h } }",
"type": "inline_equation"
},
{
"bbox": [
392,
123,
433,
138
],
"score": 1.0,
"content": "such that",
"type": "text"
}
],
"index": 3
}
],
"index": 2,
"bbox_fs": [
105,
101,
506,
138
]
},
{
"type": "interline_equation",
"bbox": [
222,
140,
388,
155
],
"lines": [
{
"bbox": [
222,
140,
388,
155
],
"spans": [
{
"bbox": [
222,
140,
388,
155
],
"score": 0.89,
"content": "\\begin{array} { r } { \\hat { f } \\left( \\mathbf { X } + \\mathbf { V } \\right) \\neq \\hat { f } \\left( \\mathbf { X } \\right) \\mathrm { f o r } ^ { * } \\mathrm { m o s t } ^ { * } \\mathbf { X } \\sim D , } \\end{array}",
"type": "interline_equation",
"image_path": "831387fcf3fae8922d58bc63c66078fe603ae4b2c440daa59c8a14c9c37671c9.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
222,
140,
388,
155
],
"spans": [],
"index": 4
}
]
},
{
"type": "text",
"bbox": [
106,
158,
507,
181
],
"lines": [
{
"bbox": [
106,
158,
505,
171
],
"spans": [
{
"bbox": [
106,
158,
132,
171
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
133,
159,
142,
168
],
"score": 0.83,
"content": "D",
"type": "inline_equation"
},
{
"bbox": [
143,
158,
505,
171
],
"score": 1.0,
"content": "is the distribution of the EEG data. The UAP can be determined by optimizing the negative",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
169,
495,
182
],
"spans": [
{
"bbox": [
106,
169,
244,
182
],
"score": 1.0,
"content": "log-likelihood loss with respect to",
"type": "text"
},
{
"bbox": [
245,
169,
255,
180
],
"score": 0.69,
"content": "\\mathbf { V }",
"type": "inline_equation"
},
{
"bbox": [
255,
169,
495,
182
],
"score": 1.0,
"content": "using batch gradient descent on the trials in the training set.",
"type": "text"
}
],
"index": 6
}
],
"index": 5.5,
"bbox_fs": [
106,
158,
505,
182
]
},
{
"type": "title",
"bbox": [
107,
196,
333,
209
],
"lines": [
{
"bbox": [
105,
195,
334,
210
],
"spans": [
{
"bbox": [
105,
195,
334,
210
],
"score": 1.0,
"content": "3 MODELING PRACTICAL ATTACKS IN BCI",
"type": "text"
}
],
"index": 7
}
],
"index": 7
},
{
"type": "text",
"bbox": [
106,
220,
506,
342
],
"lines": [
{
"bbox": [
105,
220,
507,
234
],
"spans": [
{
"bbox": [
105,
220,
507,
234
],
"score": 1.0,
"content": "This section is the main contribution of the paper: we present a design of practical DoS attacks on MI-",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
231,
505,
245
],
"spans": [
{
"bbox": [
105,
231,
505,
245
],
"score": 1.0,
"content": "BCIs that operates at the source of the signal acquisition. We propose a new method to eliminate the",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
243,
506,
255
],
"spans": [
{
"bbox": [
105,
243,
506,
255
],
"score": 1.0,
"content": "square wave artifacts to generate adversarial examples that are natural and physiologically plausible.",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
253,
505,
267
],
"spans": [
{
"bbox": [
105,
253,
505,
267
],
"score": 1.0,
"content": "The perturbation is emitted by a smart, adversarial device placed close to the ear, e.g., a smart glass or",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
265,
505,
278
],
"spans": [
{
"bbox": [
105,
265,
505,
278
],
"score": 1.0,
"content": "in-ear headphones, and is propagated to the individual EEG electrodes over the scalp’s skin. As can",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
276,
506,
288
],
"spans": [
{
"bbox": [
105,
276,
506,
288
],
"score": 1.0,
"content": "be experimentally observed on measured EEG traces (Merlet et al., 2013; Sazgar & Young, 2019),",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
287,
506,
300
],
"spans": [
{
"bbox": [
106,
287,
506,
300
],
"score": 1.0,
"content": "the same electrical source, e.g., electrocardiographic activities, is sensed by each EEG electrode with",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
298,
506,
311
],
"spans": [
{
"bbox": [
106,
298,
506,
311
],
"score": 1.0,
"content": "different degrees of attenuation and delay. We present a practical propagation model that determines",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
309,
505,
320
],
"spans": [
{
"bbox": [
106,
309,
505,
320
],
"score": 1.0,
"content": "the magnitude and delay for every individual electrode based on the distance along the scalp to the",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
320,
506,
333
],
"spans": [
{
"bbox": [
106,
320,
506,
333
],
"score": 1.0,
"content": "adversarial device. The perturbation is trained end-to-end to fool the classifier to always output “rest,”",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
330,
506,
343
],
"spans": [
{
"bbox": [
105,
330,
506,
343
],
"score": 1.0,
"content": "hence DoS, while respecting the spatial model and the amplitude constraints to remain imperceptible.",
"type": "text"
}
],
"index": 18
}
],
"index": 13,
"bbox_fs": [
105,
220,
507,
343
]
},
{
"type": "title",
"bbox": [
107,
354,
437,
367
],
"lines": [
{
"bbox": [
105,
354,
438,
368
],
"spans": [
{
"bbox": [
105,
354,
438,
368
],
"score": 1.0,
"content": "3.1 DESIGN AND ASSESSMENT OF PHYSIOLOGICALLY PLAUSIBLE ATTACKS",
"type": "text"
}
],
"index": 19
}
],
"index": 19
},
{
"type": "text",
"bbox": [
106,
375,
506,
454
],
"lines": [
{
"bbox": [
105,
375,
506,
388
],
"spans": [
{
"bbox": [
105,
375,
506,
388
],
"score": 1.0,
"content": "PGD-designed attacks on EEG tend to form perturbation signals which resemble a square-wave",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
104,
386,
507,
400
],
"spans": [
{
"bbox": [
104,
386,
507,
400
],
"score": 1.0,
"content": "artifact (see Figure 2), an effect that has been observed on ECG data, too (Han et al., 2020). However,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
397,
506,
410
],
"spans": [
{
"bbox": [
105,
397,
506,
410
],
"score": 1.0,
"content": "EEG signals are of random nature and can be modeled as frequency dependent stationary or non-",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
408,
506,
421
],
"spans": [
{
"bbox": [
105,
408,
506,
421
],
"score": 1.0,
"content": "stationary random processes (Karlekar & Gupta, 2014). To this end, we introduce a new loss term in",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
420,
505,
432
],
"spans": [
{
"bbox": [
106,
420,
505,
432
],
"score": 1.0,
"content": "the PGD optimization such that the perturbation resembles the random nature of EEG signals, which",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
430,
505,
443
],
"spans": [
{
"bbox": [
105,
430,
505,
443
],
"score": 1.0,
"content": "we achieve by promoting signal changes represented in the first order derivative. We estimate the",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
103,
439,
505,
456
],
"spans": [
{
"bbox": [
103,
439,
197,
456
],
"score": 1.0,
"content": "per-channel derivative",
"type": "text"
},
{
"bbox": [
197,
441,
365,
455
],
"score": 0.92,
"content": "\\begin{array} { r } { \\bar { { \\bf V } ^ { \\prime } } = ( { \\bf v } _ { 0 } ^ { \\prime } , { \\bf v } _ { 1 } ^ { \\prime } , . . . , { \\bf v } _ { N _ { c h } - 1 } ^ { \\prime } ) \\in \\mathbb { R } ^ { N _ { s } - 1 \\times N _ { c h } } } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
365,
439,
505,
456
],
"score": 1.0,
"content": "using the sample-wise difference:",
"type": "text"
}
],
"index": 26
}
],
"index": 23,
"bbox_fs": [
103,
375,
507,
456
]
},
{
"type": "interline_equation",
"bbox": [
155,
458,
456,
471
],
"lines": [
{
"bbox": [
155,
458,
456,
471
],
"spans": [
{
"bbox": [
155,
458,
456,
471
],
"score": 0.83,
"content": "\\begin{array} { r } { \\mathbf { v } _ { c } ^ { \\prime } [ t ] : = \\mathbf { v } _ { c } [ t ] - \\mathbf { v } _ { c } [ t - 1 ] \\quad t \\in \\{ 1 , 2 , . . . , N _ { s } - 1 \\} , c \\in \\{ 0 , 1 , . . . , N _ { c h } - 1 \\} } \\end{array}",
"type": "interline_equation",
"image_path": "ce0c71c6986adead5a7ef4f07c8039c9cf9397e84e03cd7c9175ee0546e8ce8a.jpg"
}
]
}
],
"index": 27,
"virtual_lines": [
{
"bbox": [
155,
458,
456,
471
],
"spans": [],
"index": 27
}
]
},
{
"type": "text",
"bbox": [
106,
476,
505,
515
],
"lines": [
{
"bbox": [
105,
472,
508,
496
],
"spans": [
{
"bbox": [
105,
477,
262,
492
],
"score": 1.0,
"content": "The additive loss term is determined by",
"type": "text"
},
{
"bbox": [
262,
475,
370,
491
],
"score": 0.91,
"content": "\\begin{array} { r } { l _ { 1 } ( { \\bf V } ) = - \\frac { \\beta } { \\epsilon } \\sum _ { c = 1 } ^ { N _ { c h } } | | { \\bf v } _ { c } ^ { \\prime } | | _ { 1 } } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
303,
472,
508,
496
],
"score": 1.0,
"content": "− β\u000f PNchc=1 ||v0c||1, where || · ||1 is the `1-norm, \u000f the",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
489,
505,
502
],
"spans": [
{
"bbox": [
106,
489,
260,
502
],
"score": 1.0,
"content": "maximum perturbation amplitude, and",
"type": "text"
},
{
"bbox": [
261,
490,
286,
500
],
"score": 0.9,
"content": "\\beta \\geq 0",
"type": "inline_equation"
},
{
"bbox": [
286,
489,
505,
502
],
"score": 1.0,
"content": "a weighting factor. When designing a one-dimensional",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
500,
347,
515
],
"spans": [
{
"bbox": [
105,
500,
271,
515
],
"score": 1.0,
"content": "perturbation, the derivative loss becomes",
"type": "text"
},
{
"bbox": [
272,
500,
343,
515
],
"score": 0.94,
"content": "\\begin{array} { r } { l _ { 1 } ( \\mathbf { v } ) = - \\frac { \\beta } { \\epsilon } | | \\mathbf { v } ^ { \\prime } | | } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
343,
500,
347,
515
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 30
}
],
"index": 29,
"bbox_fs": [
105,
472,
508,
515
]
},
{
"type": "text",
"bbox": [
107,
524,
505,
580
],
"lines": [
{
"bbox": [
105,
524,
505,
538
],
"spans": [
{
"bbox": [
105,
524,
505,
538
],
"score": 1.0,
"content": "Measuring the Plausibility of Attacks None of the previous works have given quantitative",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
536,
505,
548
],
"spans": [
{
"bbox": [
105,
536,
505,
548
],
"score": 1.0,
"content": "measures to assess the physiological plausibility of an EEG adversarial attack. In this work, we",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
547,
505,
559
],
"spans": [
{
"bbox": [
105,
547,
505,
559
],
"score": 1.0,
"content": "propose data-driven measures for quantifying the naturalism of an attack. We compute either the",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
558,
504,
570
],
"spans": [
{
"bbox": [
106,
558,
504,
570
],
"score": 1.0,
"content": "cross correlation, the Euclidian distance, or the cosine similarity between the attacked signal and the",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
569,
491,
581
],
"spans": [
{
"bbox": [
106,
569,
291,
581
],
"score": 1.0,
"content": "original EEG, and average the values over the",
"type": "text"
},
{
"bbox": [
291,
569,
309,
580
],
"score": 0.91,
"content": "N _ { c h }",
"type": "inline_equation"
},
{
"bbox": [
309,
569,
491,
581
],
"score": 1.0,
"content": "channels and over the samples in the dataset.",
"type": "text"
}
],
"index": 35
}
],
"index": 33,
"bbox_fs": [
105,
524,
505,
581
]
},
{
"type": "title",
"bbox": [
108,
593,
262,
604
],
"lines": [
{
"bbox": [
106,
593,
264,
606
],
"spans": [
{
"bbox": [
106,
593,
264,
606
],
"score": 1.0,
"content": "3.2 SPATIAL PROPAGATION MODEL",
"type": "text"
}
],
"index": 36
}
],
"index": 36
},
{
"type": "text",
"bbox": [
106,
613,
505,
713
],
"lines": [
{
"bbox": [
106,
613,
506,
627
],
"spans": [
{
"bbox": [
106,
613,
506,
627
],
"score": 1.0,
"content": "So far, a perturbation signal was designed for every individual channel. It is unrealistic for an attacker",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
624,
505,
637
],
"spans": [
{
"bbox": [
105,
624,
505,
637
],
"score": 1.0,
"content": "to perturb the signal for all individual channels simultaneously; hence, we consider a more practical",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
104,
633,
506,
649
],
"spans": [
{
"bbox": [
104,
633,
256,
649
],
"score": 1.0,
"content": "use case where the perturbation signal",
"type": "text"
},
{
"bbox": [
257,
635,
293,
646
],
"score": 0.91,
"content": "\\mathbf { v } \\in \\mathbb { R } ^ { N _ { s } }",
"type": "inline_equation"
},
{
"bbox": [
293,
633,
506,
649
],
"score": 1.0,
"content": "is emitted from one location, e.g., from an adversarial",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
645,
505,
659
],
"spans": [
{
"bbox": [
105,
645,
505,
659
],
"score": 1.0,
"content": "device placed on the left side of the subject or close to the left ear. More specifically, in this study, we",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
658,
506,
671
],
"spans": [
{
"bbox": [
105,
658,
506,
671
],
"score": 1.0,
"content": "assume that the EEG electrode at the position T9 according to the international 10-10 system (Sch),",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
667,
505,
683
],
"spans": [
{
"bbox": [
105,
667,
505,
683
],
"score": 1.0,
"content": "which is the closest to the left ear, senses the largest perturbation. The signal subsequently propagates",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
678,
505,
693
],
"spans": [
{
"bbox": [
105,
678,
505,
693
],
"score": 1.0,
"content": "over the skin to each electrode, which results in an individual magnitude and delay depending on",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
690,
506,
703
],
"spans": [
{
"bbox": [
106,
690,
506,
703
],
"score": 1.0,
"content": "the distance between the adversarial device and the electrode. More formally, we model the sensed",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
702,
290,
713
],
"spans": [
{
"bbox": [
105,
702,
201,
713
],
"score": 1.0,
"content": "perturbation at channel",
"type": "text"
},
{
"bbox": [
201,
702,
205,
711
],
"score": 0.75,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
206,
702,
273,
713
],
"score": 1.0,
"content": "and time instant",
"type": "text"
},
{
"bbox": [
273,
703,
278,
711
],
"score": 0.8,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
278,
702,
290,
713
],
"score": 1.0,
"content": "as",
"type": "text"
}
],
"index": 45
}
],
"index": 41,
"bbox_fs": [
104,
613,
506,
713
]
},
{
"type": "interline_equation",
"bbox": [
204,
716,
406,
730
],
"lines": [
{
"bbox": [
204,
716,
406,
730
],
"spans": [
{
"bbox": [
204,
716,
406,
730
],
"score": 0.88,
"content": "h _ { i } ( \\mathbf { v } , \\lambda _ { m } , \\lambda _ { d } ) ( t ) : = m ( l _ { i } , \\lambda _ { m } ) \\cdot \\mathbf { v } \\left( t - d ( l _ { i } , \\lambda _ { d } ) \\right) ,",
"type": "interline_equation",
"image_path": "376c0312852c9f41d7a3778ae583d8e40b0cf5932a3541026cd26602e02373b7.jpg"
}
]
}
],
"index": 46,
"virtual_lines": [
{
"bbox": [
204,
716,
406,
730
],
"spans": [],
"index": 46
}
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
116
],
"lines": [
{
"bbox": [
105,
82,
505,
95
],
"spans": [
{
"bbox": [
105,
82,
133,
95
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
82,
175,
95
],
"score": 0.92,
"content": "m ( l _ { i } , \\lambda _ { m } )",
"type": "inline_equation"
},
{
"bbox": [
175,
82,
194,
95
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
194,
82,
228,
95
],
"score": 0.94,
"content": "d ( l _ { i } , \\lambda _ { d } )",
"type": "inline_equation"
},
{
"bbox": [
229,
82,
505,
95
],
"score": 1.0,
"content": "are the magnitude and the delay respectively, both of which depend",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
94,
504,
105
],
"spans": [
{
"bbox": [
106,
94,
168,
105
],
"score": 1.0,
"content": "on the distance",
"type": "text"
},
{
"bbox": [
168,
95,
175,
104
],
"score": 0.85,
"content": "l _ { i }",
"type": "inline_equation"
},
{
"bbox": [
176,
94,
306,
105
],
"score": 1.0,
"content": "and on characteristic parameters",
"type": "text"
},
{
"bbox": [
306,
94,
321,
105
],
"score": 0.9,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
321,
94,
338,
105
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
339,
94,
350,
105
],
"score": 0.88,
"content": "\\lambda _ { d }",
"type": "inline_equation"
},
{
"bbox": [
351,
94,
504,
105
],
"score": 1.0,
"content": ". We define the resulting multi-channel",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
104,
101,
430,
118
],
"spans": [
{
"bbox": [
104,
101,
158,
118
],
"score": 1.0,
"content": "perturbation",
"type": "text"
},
{
"bbox": [
158,
104,
217,
115
],
"score": 0.91,
"content": "\\mathbf { V } \\in \\mathbb { R } ^ { N _ { s } \\times N _ { c h } }",
"type": "inline_equation"
},
{
"bbox": [
218,
101,
430,
118
],
"score": 1.0,
"content": ", which is added to the multi-channel EEG signal, as",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "interline_equation",
"bbox": [
118,
121,
476,
135
],
"lines": [
{
"bbox": [
118,
121,
476,
135
],
"spans": [
{
"bbox": [
118,
121,
476,
135
],
"score": 0.9,
"content": "\\begin{array} { r } { { \\bf V } ( \\lambda _ { m } , \\lambda _ { d } ) = H ( { \\bf v } , \\lambda _ { m } , \\lambda _ { d } ) : = \\left( h _ { 0 } ( { \\bf v } , \\lambda _ { m } , \\lambda _ { d } ) , h _ { 1 } ( { \\bf v } , \\lambda _ { m } , \\lambda _ { d } ) , . . . , h _ { N _ { c h } - 1 } ( { \\bf v } , \\lambda _ { m } , \\lambda _ { d } ) \\right) . } \\end{array}",
"type": "interline_equation",
"image_path": "de96e5a139b1fc23afd99dccaff6c949ce6e68096635c4b34caf36537f06be33.jpg"
}
]
}
],
"index": 3,
"virtual_lines": [
{
"bbox": [
118,
121,
476,
135
],
"spans": [],
"index": 3
}
]
},
{
"type": "text",
"bbox": [
106,
146,
505,
180
],
"lines": [
{
"bbox": [
106,
146,
505,
159
],
"spans": [
{
"bbox": [
106,
146,
204,
159
],
"score": 1.0,
"content": "We estimate the distance",
"type": "text"
},
{
"bbox": [
205,
147,
212,
158
],
"score": 0.86,
"content": "l _ { i }",
"type": "inline_equation"
},
{
"bbox": [
212,
146,
505,
159
],
"score": 1.0,
"content": "between the electrode at position T9 and the remaining, attacked positions",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
158,
505,
170
],
"spans": [
{
"bbox": [
106,
158,
337,
170
],
"score": 1.0,
"content": "using the 10-10 system and a head model with a radius of",
"type": "text"
},
{
"bbox": [
338,
158,
365,
168
],
"score": 0.61,
"content": "8 . 7 \\mathrm { c m }",
"type": "inline_equation"
},
{
"bbox": [
366,
158,
505,
170
],
"score": 1.0,
"content": "(Algazi et al., 2001). We decouple",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
168,
506,
182
],
"spans": [
{
"bbox": [
105,
168,
506,
182
],
"score": 1.0,
"content": "the distance-dependent modeling of the magnitude and delay, explained in the following paragraphs.",
"type": "text"
}
],
"index": 6
}
],
"index": 5
},
{
"type": "text",
"bbox": [
106,
192,
505,
236
],
"lines": [
{
"bbox": [
106,
192,
505,
204
],
"spans": [
{
"bbox": [
106,
192,
505,
204
],
"score": 1.0,
"content": "Magnitude. For modeling the magnitude, we assume that the adversarial device injects or induces",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
203,
505,
215
],
"spans": [
{
"bbox": [
105,
203,
144,
215
],
"score": 1.0,
"content": "a current",
"type": "text"
},
{
"bbox": [
144,
204,
151,
213
],
"score": 0.73,
"content": "I",
"type": "inline_equation"
},
{
"bbox": [
151,
203,
234,
215
],
"score": 1.0,
"content": ", yielding a potential",
"type": "text"
},
{
"bbox": [
235,
204,
244,
213
],
"score": 0.79,
"content": "V",
"type": "inline_equation"
},
{
"bbox": [
244,
203,
505,
215
],
"score": 1.0,
"content": "measured near T9. The current propagates over the head surface",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
214,
505,
227
],
"spans": [
{
"bbox": [
106,
214,
505,
227
],
"score": 1.0,
"content": "through the skin to each of the remaining attacked electrodes, which can be modeled as a cylindrical",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
224,
201,
237
],
"spans": [
{
"bbox": [
105,
224,
201,
237
],
"score": 1.0,
"content": "resistor with resistance",
"type": "text"
}
],
"index": 10
}
],
"index": 8.5
},
{
"type": "interline_equation",
"bbox": [
282,
239,
327,
264
],
"lines": [
{
"bbox": [
282,
239,
327,
264
],
"spans": [
{
"bbox": [
282,
239,
327,
264
],
"score": 0.93,
"content": "R _ { i } = \\frac { l _ { i } } { \\sigma A } ,",
"type": "interline_equation",
"image_path": "df7117bdb4d5c93cccf2e90b25cc3c804499313daf11d78a2ed1cb0219bdb070.jpg"
}
]
}
],
"index": 11,
"virtual_lines": [
{
"bbox": [
282,
239,
327,
264
],
"spans": [],
"index": 11
}
]
},
{
"type": "text",
"bbox": [
107,
268,
505,
302
],
"lines": [
{
"bbox": [
106,
268,
506,
280
],
"spans": [
{
"bbox": [
106,
268,
132,
280
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
132,
271,
140,
279
],
"score": 0.78,
"content": "\\sigma",
"type": "inline_equation"
},
{
"bbox": [
140,
268,
506,
280
],
"score": 1.0,
"content": "is the conductivity of the skin which can be in the range of [0.28, 0.87] Siemens/m (Vorwerk",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
279,
507,
293
],
"spans": [
{
"bbox": [
105,
279,
176,
293
],
"score": 1.0,
"content": "et al., 2019), and",
"type": "text"
},
{
"bbox": [
176,
280,
184,
290
],
"score": 0.81,
"content": "A",
"type": "inline_equation"
},
{
"bbox": [
185,
279,
420,
293
],
"score": 1.0,
"content": "is the area of the skin conductor. The potential at electrode",
"type": "text"
},
{
"bbox": [
420,
281,
425,
290
],
"score": 0.77,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
425,
279,
435,
293
],
"score": 1.0,
"content": "is",
"type": "text"
},
{
"bbox": [
436,
280,
502,
291
],
"score": 0.92,
"content": "V _ { i } = V - I \\cdot R _ { i }",
"type": "inline_equation"
},
{
"bbox": [
502,
279,
507,
293
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
291,
288,
303
],
"spans": [
{
"bbox": [
106,
291,
288,
303
],
"score": 1.0,
"content": "and hence the magnitude can be described as",
"type": "text"
}
],
"index": 14
}
],
"index": 13
},
{
"type": "interline_equation",
"bbox": [
196,
307,
414,
332
],
"lines": [
{
"bbox": [
196,
307,
414,
332
],
"spans": [
{
"bbox": [
196,
307,
414,
332
],
"score": 0.92,
"content": "m ( l _ { i } , \\lambda _ { m } ) = 1 - \\frac { V - V _ { i } } { V } = 1 - \\frac { I } { V \\sigma A } l _ { i } = 1 - \\lambda _ { m } l _ { i } ,",
"type": "interline_equation",
"image_path": "be06049f08f97d940ab4e03b5c05093432b85058b4a81d8046d097f375f7225f.jpg"
}
]
}
],
"index": 15,
"virtual_lines": [
{
"bbox": [
196,
307,
414,
332
],
"spans": [],
"index": 15
}
]
},
{
"type": "text",
"bbox": [
106,
335,
506,
426
],
"lines": [
{
"bbox": [
105,
335,
504,
350
],
"spans": [
{
"bbox": [
105,
335,
223,
350
],
"score": 1.0,
"content": "where we further constrain",
"type": "text"
},
{
"bbox": [
223,
336,
312,
349
],
"score": 0.92,
"content": "0 \\leq m ( l _ { i } , \\lambda _ { m } ) \\leq 1",
"type": "inline_equation"
},
{
"bbox": [
312,
335,
489,
350
],
"score": 1.0,
"content": ". The characteristic magnitude parameter",
"type": "text"
},
{
"bbox": [
489,
337,
504,
348
],
"score": 0.89,
"content": "\\lambda _ { m }",
"type": "inline_equation"
}
],
"index": 16
},
{
"bbox": [
104,
347,
506,
362
],
"spans": [
{
"bbox": [
104,
347,
506,
362
],
"score": 1.0,
"content": "represents the complex interplay between input current, voltage, conductivity, and area, covering",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
104,
358,
506,
371
],
"spans": [
{
"bbox": [
104,
358,
450,
371
],
"score": 1.0,
"content": "various attack scenarios. We consider different characteristic magnitude parameters",
"type": "text"
},
{
"bbox": [
450,
358,
502,
370
],
"score": 0.79,
"content": "\\lambda _ { m } \\in [ 1 , 1 5 ]",
"type": "inline_equation"
},
{
"bbox": [
503,
358,
506,
371
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
369,
506,
383
],
"spans": [
{
"bbox": [
105,
369,
140,
383
],
"score": 1.0,
"content": "A large",
"type": "text"
},
{
"bbox": [
140,
370,
154,
381
],
"score": 0.89,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
155,
369,
506,
383
],
"score": 1.0,
"content": "represents cases with large attenuation and limited propagation, i.e., a limited set of",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
106,
381,
506,
393
],
"spans": [
{
"bbox": [
106,
381,
387,
393
],
"score": 1.0,
"content": "neighboring electrodes sense the perturbation. Conversely, a small",
"type": "text"
},
{
"bbox": [
387,
381,
402,
392
],
"score": 0.9,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
402,
381,
506,
393
],
"score": 1.0,
"content": "covers cases with lower",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
392,
506,
404
],
"spans": [
{
"bbox": [
105,
392,
506,
404
],
"score": 1.0,
"content": "attenuation where the perturbation can propagate further and infects all electrodes. We consider also",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
402,
507,
415
],
"spans": [
{
"bbox": [
105,
402,
468,
415
],
"score": 1.0,
"content": "an intermediate case where around half of the electrodes are affected by the attack with",
"type": "text"
},
{
"bbox": [
469,
403,
503,
414
],
"score": 0.91,
"content": "\\lambda _ { m } = 5",
"type": "inline_equation"
},
{
"bbox": [
503,
402,
507,
415
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
414,
486,
426
],
"spans": [
{
"bbox": [
105,
414,
486,
426
],
"score": 1.0,
"content": "Appendix B provides examples of the magnitude of the spatial propagation on the head model.",
"type": "text"
}
],
"index": 23
}
],
"index": 19.5
},
{
"type": "text",
"bbox": [
106,
437,
505,
537
],
"lines": [
{
"bbox": [
105,
437,
506,
450
],
"spans": [
{
"bbox": [
105,
437,
506,
450
],
"score": 1.0,
"content": "Delay. The propagation of a signal on the head surface yields a position-dependent phase angle or",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
448,
505,
461
],
"spans": [
{
"bbox": [
106,
448,
505,
461
],
"score": 1.0,
"content": "delay, as shown by experimental measurements of related studies (Plutchik & Hirsch, 1963; Qiao",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
460,
506,
471
],
"spans": [
{
"bbox": [
106,
460,
506,
471
],
"score": 1.0,
"content": "et al., 1994). The delay stems from a combination of resistive and capacitive components that are",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
470,
506,
482
],
"spans": [
{
"bbox": [
105,
470,
506,
482
],
"score": 1.0,
"content": "encountered during the propagation of the signal, which can be modeled as an RC-circuit with",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
481,
506,
493
],
"spans": [
{
"bbox": [
105,
482,
149,
493
],
"score": 1.0,
"content": "resistance",
"type": "text"
},
{
"bbox": [
149,
482,
158,
491
],
"score": 0.77,
"content": "R",
"type": "inline_equation"
},
{
"bbox": [
158,
482,
198,
493
],
"score": 1.0,
"content": ", capacity",
"type": "text"
},
{
"bbox": [
198,
482,
207,
491
],
"score": 0.79,
"content": "C",
"type": "inline_equation"
},
{
"bbox": [
207,
482,
285,
493
],
"score": 1.0,
"content": ", and time constant",
"type": "text"
},
{
"bbox": [
285,
481,
327,
491
],
"score": 0.91,
"content": "\\tau = R \\cdot C",
"type": "inline_equation"
},
{
"bbox": [
328,
482,
506,
493
],
"score": 1.0,
"content": "that relates to the group delay. Specifically,",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
492,
506,
505
],
"spans": [
{
"bbox": [
106,
492,
506,
505
],
"score": 1.0,
"content": "the contacts between the electrodes and the skin are predominantly capacitive whereas the skin itself",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
104,
502,
505,
517
],
"spans": [
{
"bbox": [
104,
502,
505,
517
],
"score": 1.0,
"content": "is both resistive and capacitive (Kim et al., 2010). As explained in the previous part, an increasing",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
513,
505,
527
],
"spans": [
{
"bbox": [
105,
513,
428,
527
],
"score": 1.0,
"content": "distance between the attacker and the target electrode yields a larger resistance",
"type": "text"
},
{
"bbox": [
428,
515,
437,
524
],
"score": 0.57,
"content": "R",
"type": "inline_equation"
},
{
"bbox": [
437,
513,
505,
527
],
"score": 1.0,
"content": ". As a result, the",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
525,
279,
537
],
"spans": [
{
"bbox": [
106,
525,
162,
537
],
"score": 1.0,
"content": "time constant",
"type": "text"
},
{
"bbox": [
162,
527,
169,
535
],
"score": 0.76,
"content": "\\tau",
"type": "inline_equation"
},
{
"bbox": [
170,
525,
279,
537
],
"score": 1.0,
"content": "and the delay increase too.",
"type": "text"
}
],
"index": 32
}
],
"index": 28
},
{
"type": "text",
"bbox": [
106,
541,
506,
619
],
"lines": [
{
"bbox": [
106,
542,
506,
554
],
"spans": [
{
"bbox": [
106,
542,
506,
554
],
"score": 1.0,
"content": "Here, we model a linear distance-delay relation. We rely on a study by Plutchik & Hirsch (1963),",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
552,
506,
566
],
"spans": [
{
"bbox": [
105,
552,
506,
566
],
"score": 1.0,
"content": "which conducted human skin impedance and phase angle measurements by placing electrodes at",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
564,
506,
577
],
"spans": [
{
"bbox": [
105,
564,
219,
577
],
"score": 1.0,
"content": "an approximate distance of",
"type": "text"
},
{
"bbox": [
219,
564,
245,
574
],
"score": 0.68,
"content": "1 0 \\mathrm { c m }",
"type": "inline_equation"
},
{
"bbox": [
245,
564,
457,
577
],
"score": 1.0,
"content": "and applying voltages with frequencies in the range",
"type": "text"
},
{
"bbox": [
458,
564,
503,
574
],
"score": 0.78,
"content": "2 { \\mathrm { - } } 1 0 0 0 \\mathrm { { H z } }",
"type": "inline_equation"
},
{
"bbox": [
504,
564,
506,
577
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
575,
506,
588
],
"spans": [
{
"bbox": [
105,
575,
506,
588
],
"score": 1.0,
"content": "When assuming a linear frequency-phase relation in low-frequency region (Qiao et al., 1994), one",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
586,
505,
598
],
"spans": [
{
"bbox": [
106,
586,
234,
598
],
"score": 1.0,
"content": "can derive the group delay to be",
"type": "text"
},
{
"bbox": [
234,
586,
261,
596
],
"score": 0.43,
"content": "2 . 8 \\mathrm { m s }",
"type": "inline_equation"
},
{
"bbox": [
261,
586,
414,
598
],
"score": 1.0,
"content": "when considering a measured angle of",
"type": "text"
},
{
"bbox": [
414,
586,
429,
596
],
"score": 0.88,
"content": "1 0 ^ { \\circ }",
"type": "inline_equation"
},
{
"bbox": [
430,
586,
440,
598
],
"score": 1.0,
"content": "at",
"type": "text"
},
{
"bbox": [
441,
586,
464,
596
],
"score": 0.72,
"content": "1 0 \\mathrm { H z }",
"type": "inline_equation"
},
{
"bbox": [
465,
586,
505,
598
],
"score": 1.0,
"content": ". As those",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
104,
596,
506,
611
],
"spans": [
{
"bbox": [
104,
596,
506,
611
],
"score": 1.0,
"content": "measurements were conducted for only one distance, we extrapolate the delay for the remaining",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
609,
267,
620
],
"spans": [
{
"bbox": [
106,
609,
267,
620
],
"score": 1.0,
"content": "distances using a rectified linear model:",
"type": "text"
}
],
"index": 39
}
],
"index": 36
},
{
"type": "interline_equation",
"bbox": [
171,
624,
438,
639
],
"lines": [
{
"bbox": [
171,
624,
438,
639
],
"spans": [
{
"bbox": [
171,
624,
438,
639
],
"score": 0.89,
"content": "\\lambda _ { d } \\cdot ( l _ { i } - l _ { 0 } ) > 0 \\uparrow d ( l _ { i } , \\lambda _ { d } ) = \\lambda _ { d } \\cdot ( l _ { i } - l _ { 0 } ) + d _ { 0 } : d ( l _ { i } , \\lambda _ { d } ) = 0 ,",
"type": "interline_equation",
"image_path": "8c109f14baba8fdb3fa4ca00a590e185e17699666a46300fe54ae3c09558dc5d.jpg"
}
]
}
],
"index": 40,
"virtual_lines": [
{
"bbox": [
171,
624,
438,
639
],
"spans": [],
"index": 40
}
]
},
{
"type": "text",
"bbox": [
106,
643,
506,
732
],
"lines": [
{
"bbox": [
105,
643,
507,
657
],
"spans": [
{
"bbox": [
105,
643,
133,
657
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
133,
644,
183,
655
],
"score": 0.9,
"content": "d _ { 0 } = 2 . 8 \\mathrm { m s }",
"type": "inline_equation"
},
{
"bbox": [
184,
643,
277,
657
],
"score": 1.0,
"content": "is the delay at distance",
"type": "text"
},
{
"bbox": [
278,
644,
324,
655
],
"score": 0.93,
"content": "l _ { 0 } = 1 0 \\mathrm { c m }",
"type": "inline_equation"
},
{
"bbox": [
324,
643,
507,
657
],
"score": 1.0,
"content": ". The delay depends not only on the distance,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
655,
505,
667
],
"spans": [
{
"bbox": [
106,
655,
505,
667
],
"score": 1.0,
"content": "but also on other parameters such as the electrode-to-skin contact, the humidity of the skin, etc. To",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
665,
506,
678
],
"spans": [
{
"bbox": [
105,
665,
506,
678
],
"score": 1.0,
"content": "this end, we evaluate the propagation of the attack with different characteristic delay parameters",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
677,
506,
689
],
"spans": [
{
"bbox": [
106,
677,
192,
689
],
"score": 0.88,
"content": "\\lambda _ { d } \\in \\left[ 0 . 1 , 0 . 5 6 3 \\right] \\mathrm { s / m }",
"type": "inline_equation"
},
{
"bbox": [
192,
677,
218,
689
],
"score": 1.0,
"content": ". With",
"type": "text"
},
{
"bbox": [
218,
677,
256,
688
],
"score": 0.93,
"content": "\\lambda _ { d } = 0 . 1",
"type": "inline_equation"
},
{
"bbox": [
257,
677,
506,
689
],
"score": 1.0,
"content": "we cover the cases where very little delay happens, while the",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
688,
506,
700
],
"spans": [
{
"bbox": [
105,
688,
183,
700
],
"score": 1.0,
"content": "largest considered",
"type": "text"
},
{
"bbox": [
183,
688,
249,
699
],
"score": 0.9,
"content": "\\lambda _ { d } = 0 . 5 6 3 \\mathrm { s / m }",
"type": "inline_equation"
},
{
"bbox": [
249,
688,
365,
700
],
"score": 1.0,
"content": "yields a maximum delay of",
"type": "text"
},
{
"bbox": [
365,
688,
385,
698
],
"score": 0.34,
"content": "0 . 1 \\mathrm { s }",
"type": "inline_equation"
},
{
"bbox": [
385,
688,
506,
700
],
"score": 1.0,
"content": "at the farthest electrode T10,",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
104,
698,
507,
712
],
"spans": [
{
"bbox": [
104,
698,
507,
712
],
"score": 1.0,
"content": "which is in alignment with the observed EEG measurements (Merlet et al., 2013; Sazgar & Young,",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
106,
710,
505,
722
],
"spans": [
{
"bbox": [
106,
710,
185,
722
],
"score": 1.0,
"content": "2019). Similarly to",
"type": "text"
},
{
"bbox": [
185,
710,
199,
721
],
"score": 0.88,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
199,
710,
389,
722
],
"score": 1.0,
"content": ", we showcase also for an intermediate value of",
"type": "text"
},
{
"bbox": [
389,
710,
427,
721
],
"score": 0.91,
"content": "\\lambda _ { d } = 0 . 3",
"type": "inline_equation"
},
{
"bbox": [
427,
710,
505,
722
],
"score": 1.0,
"content": "which corresponds",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
106,
721,
228,
733
],
"spans": [
{
"bbox": [
106,
721,
158,
733
],
"score": 1.0,
"content": "to a delay of",
"type": "text"
},
{
"bbox": [
159,
721,
195,
731
],
"score": 0.27,
"content": "0 . 0 5 3 \\mathrm { m s }",
"type": "inline_equation"
},
{
"bbox": [
196,
721,
228,
733
],
"score": 1.0,
"content": "at T10.",
"type": "text"
}
],
"index": 48
}
],
"index": 44.5
}
],
"page_idx": 4,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
26,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
309,
38
],
"spans": [
{
"bbox": [
106,
25,
309,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
308,
760
],
"lines": [
{
"bbox": [
302,
750,
309,
763
],
"spans": [
{
"bbox": [
302,
750,
309,
763
],
"score": 1.0,
"content": "5",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
116
],
"lines": [
{
"bbox": [
105,
82,
505,
95
],
"spans": [
{
"bbox": [
105,
82,
133,
95
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
82,
175,
95
],
"score": 0.92,
"content": "m ( l _ { i } , \\lambda _ { m } )",
"type": "inline_equation"
},
{
"bbox": [
175,
82,
194,
95
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
194,
82,
228,
95
],
"score": 0.94,
"content": "d ( l _ { i } , \\lambda _ { d } )",
"type": "inline_equation"
},
{
"bbox": [
229,
82,
505,
95
],
"score": 1.0,
"content": "are the magnitude and the delay respectively, both of which depend",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
94,
504,
105
],
"spans": [
{
"bbox": [
106,
94,
168,
105
],
"score": 1.0,
"content": "on the distance",
"type": "text"
},
{
"bbox": [
168,
95,
175,
104
],
"score": 0.85,
"content": "l _ { i }",
"type": "inline_equation"
},
{
"bbox": [
176,
94,
306,
105
],
"score": 1.0,
"content": "and on characteristic parameters",
"type": "text"
},
{
"bbox": [
306,
94,
321,
105
],
"score": 0.9,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
321,
94,
338,
105
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
339,
94,
350,
105
],
"score": 0.88,
"content": "\\lambda _ { d }",
"type": "inline_equation"
},
{
"bbox": [
351,
94,
504,
105
],
"score": 1.0,
"content": ". We define the resulting multi-channel",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
104,
101,
430,
118
],
"spans": [
{
"bbox": [
104,
101,
158,
118
],
"score": 1.0,
"content": "perturbation",
"type": "text"
},
{
"bbox": [
158,
104,
217,
115
],
"score": 0.91,
"content": "\\mathbf { V } \\in \\mathbb { R } ^ { N _ { s } \\times N _ { c h } }",
"type": "inline_equation"
},
{
"bbox": [
218,
101,
430,
118
],
"score": 1.0,
"content": ", which is added to the multi-channel EEG signal, as",
"type": "text"
}
],
"index": 2
}
],
"index": 1,
"bbox_fs": [
104,
82,
505,
118
]
},
{
"type": "interline_equation",
"bbox": [
118,
121,
476,
135
],
"lines": [
{
"bbox": [
118,
121,
476,
135
],
"spans": [
{
"bbox": [
118,
121,
476,
135
],
"score": 0.9,
"content": "\\begin{array} { r } { { \\bf V } ( \\lambda _ { m } , \\lambda _ { d } ) = H ( { \\bf v } , \\lambda _ { m } , \\lambda _ { d } ) : = \\left( h _ { 0 } ( { \\bf v } , \\lambda _ { m } , \\lambda _ { d } ) , h _ { 1 } ( { \\bf v } , \\lambda _ { m } , \\lambda _ { d } ) , . . . , h _ { N _ { c h } - 1 } ( { \\bf v } , \\lambda _ { m } , \\lambda _ { d } ) \\right) . } \\end{array}",
"type": "interline_equation",
"image_path": "de96e5a139b1fc23afd99dccaff6c949ce6e68096635c4b34caf36537f06be33.jpg"
}
]
}
],
"index": 3,
"virtual_lines": [
{
"bbox": [
118,
121,
476,
135
],
"spans": [],
"index": 3
}
]
},
{
"type": "text",
"bbox": [
106,
146,
505,
180
],
"lines": [
{
"bbox": [
106,
146,
505,
159
],
"spans": [
{
"bbox": [
106,
146,
204,
159
],
"score": 1.0,
"content": "We estimate the distance",
"type": "text"
},
{
"bbox": [
205,
147,
212,
158
],
"score": 0.86,
"content": "l _ { i }",
"type": "inline_equation"
},
{
"bbox": [
212,
146,
505,
159
],
"score": 1.0,
"content": "between the electrode at position T9 and the remaining, attacked positions",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
158,
505,
170
],
"spans": [
{
"bbox": [
106,
158,
337,
170
],
"score": 1.0,
"content": "using the 10-10 system and a head model with a radius of",
"type": "text"
},
{
"bbox": [
338,
158,
365,
168
],
"score": 0.61,
"content": "8 . 7 \\mathrm { c m }",
"type": "inline_equation"
},
{
"bbox": [
366,
158,
505,
170
],
"score": 1.0,
"content": "(Algazi et al., 2001). We decouple",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
168,
506,
182
],
"spans": [
{
"bbox": [
105,
168,
506,
182
],
"score": 1.0,
"content": "the distance-dependent modeling of the magnitude and delay, explained in the following paragraphs.",
"type": "text"
}
],
"index": 6
}
],
"index": 5,
"bbox_fs": [
105,
146,
506,
182
]
},
{
"type": "text",
"bbox": [
106,
192,
505,
236
],
"lines": [
{
"bbox": [
106,
192,
505,
204
],
"spans": [
{
"bbox": [
106,
192,
505,
204
],
"score": 1.0,
"content": "Magnitude. For modeling the magnitude, we assume that the adversarial device injects or induces",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
203,
505,
215
],
"spans": [
{
"bbox": [
105,
203,
144,
215
],
"score": 1.0,
"content": "a current",
"type": "text"
},
{
"bbox": [
144,
204,
151,
213
],
"score": 0.73,
"content": "I",
"type": "inline_equation"
},
{
"bbox": [
151,
203,
234,
215
],
"score": 1.0,
"content": ", yielding a potential",
"type": "text"
},
{
"bbox": [
235,
204,
244,
213
],
"score": 0.79,
"content": "V",
"type": "inline_equation"
},
{
"bbox": [
244,
203,
505,
215
],
"score": 1.0,
"content": "measured near T9. The current propagates over the head surface",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
214,
505,
227
],
"spans": [
{
"bbox": [
106,
214,
505,
227
],
"score": 1.0,
"content": "through the skin to each of the remaining attacked electrodes, which can be modeled as a cylindrical",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
224,
201,
237
],
"spans": [
{
"bbox": [
105,
224,
201,
237
],
"score": 1.0,
"content": "resistor with resistance",
"type": "text"
}
],
"index": 10
}
],
"index": 8.5,
"bbox_fs": [
105,
192,
505,
237
]
},
{
"type": "interline_equation",
"bbox": [
282,
239,
327,
264
],
"lines": [
{
"bbox": [
282,
239,
327,
264
],
"spans": [
{
"bbox": [
282,
239,
327,
264
],
"score": 0.93,
"content": "R _ { i } = \\frac { l _ { i } } { \\sigma A } ,",
"type": "interline_equation",
"image_path": "df7117bdb4d5c93cccf2e90b25cc3c804499313daf11d78a2ed1cb0219bdb070.jpg"
}
]
}
],
"index": 11,
"virtual_lines": [
{
"bbox": [
282,
239,
327,
264
],
"spans": [],
"index": 11
}
]
},
{
"type": "text",
"bbox": [
107,
268,
505,
302
],
"lines": [
{
"bbox": [
106,
268,
506,
280
],
"spans": [
{
"bbox": [
106,
268,
132,
280
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
132,
271,
140,
279
],
"score": 0.78,
"content": "\\sigma",
"type": "inline_equation"
},
{
"bbox": [
140,
268,
506,
280
],
"score": 1.0,
"content": "is the conductivity of the skin which can be in the range of [0.28, 0.87] Siemens/m (Vorwerk",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
279,
507,
293
],
"spans": [
{
"bbox": [
105,
279,
176,
293
],
"score": 1.0,
"content": "et al., 2019), and",
"type": "text"
},
{
"bbox": [
176,
280,
184,
290
],
"score": 0.81,
"content": "A",
"type": "inline_equation"
},
{
"bbox": [
185,
279,
420,
293
],
"score": 1.0,
"content": "is the area of the skin conductor. The potential at electrode",
"type": "text"
},
{
"bbox": [
420,
281,
425,
290
],
"score": 0.77,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
425,
279,
435,
293
],
"score": 1.0,
"content": "is",
"type": "text"
},
{
"bbox": [
436,
280,
502,
291
],
"score": 0.92,
"content": "V _ { i } = V - I \\cdot R _ { i }",
"type": "inline_equation"
},
{
"bbox": [
502,
279,
507,
293
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
291,
288,
303
],
"spans": [
{
"bbox": [
106,
291,
288,
303
],
"score": 1.0,
"content": "and hence the magnitude can be described as",
"type": "text"
}
],
"index": 14
}
],
"index": 13,
"bbox_fs": [
105,
268,
507,
303
]
},
{
"type": "interline_equation",
"bbox": [
196,
307,
414,
332
],
"lines": [
{
"bbox": [
196,
307,
414,
332
],
"spans": [
{
"bbox": [
196,
307,
414,
332
],
"score": 0.92,
"content": "m ( l _ { i } , \\lambda _ { m } ) = 1 - \\frac { V - V _ { i } } { V } = 1 - \\frac { I } { V \\sigma A } l _ { i } = 1 - \\lambda _ { m } l _ { i } ,",
"type": "interline_equation",
"image_path": "be06049f08f97d940ab4e03b5c05093432b85058b4a81d8046d097f375f7225f.jpg"
}
]
}
],
"index": 15,
"virtual_lines": [
{
"bbox": [
196,
307,
414,
332
],
"spans": [],
"index": 15
}
]
},
{
"type": "text",
"bbox": [
106,
335,
506,
426
],
"lines": [
{
"bbox": [
105,
335,
504,
350
],
"spans": [
{
"bbox": [
105,
335,
223,
350
],
"score": 1.0,
"content": "where we further constrain",
"type": "text"
},
{
"bbox": [
223,
336,
312,
349
],
"score": 0.92,
"content": "0 \\leq m ( l _ { i } , \\lambda _ { m } ) \\leq 1",
"type": "inline_equation"
},
{
"bbox": [
312,
335,
489,
350
],
"score": 1.0,
"content": ". The characteristic magnitude parameter",
"type": "text"
},
{
"bbox": [
489,
337,
504,
348
],
"score": 0.89,
"content": "\\lambda _ { m }",
"type": "inline_equation"
}
],
"index": 16
},
{
"bbox": [
104,
347,
506,
362
],
"spans": [
{
"bbox": [
104,
347,
506,
362
],
"score": 1.0,
"content": "represents the complex interplay between input current, voltage, conductivity, and area, covering",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
104,
358,
506,
371
],
"spans": [
{
"bbox": [
104,
358,
450,
371
],
"score": 1.0,
"content": "various attack scenarios. We consider different characteristic magnitude parameters",
"type": "text"
},
{
"bbox": [
450,
358,
502,
370
],
"score": 0.79,
"content": "\\lambda _ { m } \\in [ 1 , 1 5 ]",
"type": "inline_equation"
},
{
"bbox": [
503,
358,
506,
371
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
369,
506,
383
],
"spans": [
{
"bbox": [
105,
369,
140,
383
],
"score": 1.0,
"content": "A large",
"type": "text"
},
{
"bbox": [
140,
370,
154,
381
],
"score": 0.89,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
155,
369,
506,
383
],
"score": 1.0,
"content": "represents cases with large attenuation and limited propagation, i.e., a limited set of",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
106,
381,
506,
393
],
"spans": [
{
"bbox": [
106,
381,
387,
393
],
"score": 1.0,
"content": "neighboring electrodes sense the perturbation. Conversely, a small",
"type": "text"
},
{
"bbox": [
387,
381,
402,
392
],
"score": 0.9,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
402,
381,
506,
393
],
"score": 1.0,
"content": "covers cases with lower",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
392,
506,
404
],
"spans": [
{
"bbox": [
105,
392,
506,
404
],
"score": 1.0,
"content": "attenuation where the perturbation can propagate further and infects all electrodes. We consider also",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
402,
507,
415
],
"spans": [
{
"bbox": [
105,
402,
468,
415
],
"score": 1.0,
"content": "an intermediate case where around half of the electrodes are affected by the attack with",
"type": "text"
},
{
"bbox": [
469,
403,
503,
414
],
"score": 0.91,
"content": "\\lambda _ { m } = 5",
"type": "inline_equation"
},
{
"bbox": [
503,
402,
507,
415
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
414,
486,
426
],
"spans": [
{
"bbox": [
105,
414,
486,
426
],
"score": 1.0,
"content": "Appendix B provides examples of the magnitude of the spatial propagation on the head model.",
"type": "text"
}
],
"index": 23
}
],
"index": 19.5,
"bbox_fs": [
104,
335,
507,
426
]
},
{
"type": "text",
"bbox": [
106,
437,
505,
537
],
"lines": [
{
"bbox": [
105,
437,
506,
450
],
"spans": [
{
"bbox": [
105,
437,
506,
450
],
"score": 1.0,
"content": "Delay. The propagation of a signal on the head surface yields a position-dependent phase angle or",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
448,
505,
461
],
"spans": [
{
"bbox": [
106,
448,
505,
461
],
"score": 1.0,
"content": "delay, as shown by experimental measurements of related studies (Plutchik & Hirsch, 1963; Qiao",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
460,
506,
471
],
"spans": [
{
"bbox": [
106,
460,
506,
471
],
"score": 1.0,
"content": "et al., 1994). The delay stems from a combination of resistive and capacitive components that are",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
470,
506,
482
],
"spans": [
{
"bbox": [
105,
470,
506,
482
],
"score": 1.0,
"content": "encountered during the propagation of the signal, which can be modeled as an RC-circuit with",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
481,
506,
493
],
"spans": [
{
"bbox": [
105,
482,
149,
493
],
"score": 1.0,
"content": "resistance",
"type": "text"
},
{
"bbox": [
149,
482,
158,
491
],
"score": 0.77,
"content": "R",
"type": "inline_equation"
},
{
"bbox": [
158,
482,
198,
493
],
"score": 1.0,
"content": ", capacity",
"type": "text"
},
{
"bbox": [
198,
482,
207,
491
],
"score": 0.79,
"content": "C",
"type": "inline_equation"
},
{
"bbox": [
207,
482,
285,
493
],
"score": 1.0,
"content": ", and time constant",
"type": "text"
},
{
"bbox": [
285,
481,
327,
491
],
"score": 0.91,
"content": "\\tau = R \\cdot C",
"type": "inline_equation"
},
{
"bbox": [
328,
482,
506,
493
],
"score": 1.0,
"content": "that relates to the group delay. Specifically,",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
492,
506,
505
],
"spans": [
{
"bbox": [
106,
492,
506,
505
],
"score": 1.0,
"content": "the contacts between the electrodes and the skin are predominantly capacitive whereas the skin itself",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
104,
502,
505,
517
],
"spans": [
{
"bbox": [
104,
502,
505,
517
],
"score": 1.0,
"content": "is both resistive and capacitive (Kim et al., 2010). As explained in the previous part, an increasing",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
513,
505,
527
],
"spans": [
{
"bbox": [
105,
513,
428,
527
],
"score": 1.0,
"content": "distance between the attacker and the target electrode yields a larger resistance",
"type": "text"
},
{
"bbox": [
428,
515,
437,
524
],
"score": 0.57,
"content": "R",
"type": "inline_equation"
},
{
"bbox": [
437,
513,
505,
527
],
"score": 1.0,
"content": ". As a result, the",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
525,
279,
537
],
"spans": [
{
"bbox": [
106,
525,
162,
537
],
"score": 1.0,
"content": "time constant",
"type": "text"
},
{
"bbox": [
162,
527,
169,
535
],
"score": 0.76,
"content": "\\tau",
"type": "inline_equation"
},
{
"bbox": [
170,
525,
279,
537
],
"score": 1.0,
"content": "and the delay increase too.",
"type": "text"
}
],
"index": 32
}
],
"index": 28,
"bbox_fs": [
104,
437,
506,
537
]
},
{
"type": "text",
"bbox": [
106,
541,
506,
619
],
"lines": [
{
"bbox": [
106,
542,
506,
554
],
"spans": [
{
"bbox": [
106,
542,
506,
554
],
"score": 1.0,
"content": "Here, we model a linear distance-delay relation. We rely on a study by Plutchik & Hirsch (1963),",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
552,
506,
566
],
"spans": [
{
"bbox": [
105,
552,
506,
566
],
"score": 1.0,
"content": "which conducted human skin impedance and phase angle measurements by placing electrodes at",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
564,
506,
577
],
"spans": [
{
"bbox": [
105,
564,
219,
577
],
"score": 1.0,
"content": "an approximate distance of",
"type": "text"
},
{
"bbox": [
219,
564,
245,
574
],
"score": 0.68,
"content": "1 0 \\mathrm { c m }",
"type": "inline_equation"
},
{
"bbox": [
245,
564,
457,
577
],
"score": 1.0,
"content": "and applying voltages with frequencies in the range",
"type": "text"
},
{
"bbox": [
458,
564,
503,
574
],
"score": 0.78,
"content": "2 { \\mathrm { - } } 1 0 0 0 \\mathrm { { H z } }",
"type": "inline_equation"
},
{
"bbox": [
504,
564,
506,
577
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
575,
506,
588
],
"spans": [
{
"bbox": [
105,
575,
506,
588
],
"score": 1.0,
"content": "When assuming a linear frequency-phase relation in low-frequency region (Qiao et al., 1994), one",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
586,
505,
598
],
"spans": [
{
"bbox": [
106,
586,
234,
598
],
"score": 1.0,
"content": "can derive the group delay to be",
"type": "text"
},
{
"bbox": [
234,
586,
261,
596
],
"score": 0.43,
"content": "2 . 8 \\mathrm { m s }",
"type": "inline_equation"
},
{
"bbox": [
261,
586,
414,
598
],
"score": 1.0,
"content": "when considering a measured angle of",
"type": "text"
},
{
"bbox": [
414,
586,
429,
596
],
"score": 0.88,
"content": "1 0 ^ { \\circ }",
"type": "inline_equation"
},
{
"bbox": [
430,
586,
440,
598
],
"score": 1.0,
"content": "at",
"type": "text"
},
{
"bbox": [
441,
586,
464,
596
],
"score": 0.72,
"content": "1 0 \\mathrm { H z }",
"type": "inline_equation"
},
{
"bbox": [
465,
586,
505,
598
],
"score": 1.0,
"content": ". As those",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
104,
596,
506,
611
],
"spans": [
{
"bbox": [
104,
596,
506,
611
],
"score": 1.0,
"content": "measurements were conducted for only one distance, we extrapolate the delay for the remaining",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
609,
267,
620
],
"spans": [
{
"bbox": [
106,
609,
267,
620
],
"score": 1.0,
"content": "distances using a rectified linear model:",
"type": "text"
}
],
"index": 39
}
],
"index": 36,
"bbox_fs": [
104,
542,
506,
620
]
},
{
"type": "interline_equation",
"bbox": [
171,
624,
438,
639
],
"lines": [
{
"bbox": [
171,
624,
438,
639
],
"spans": [
{
"bbox": [
171,
624,
438,
639
],
"score": 0.89,
"content": "\\lambda _ { d } \\cdot ( l _ { i } - l _ { 0 } ) > 0 \\uparrow d ( l _ { i } , \\lambda _ { d } ) = \\lambda _ { d } \\cdot ( l _ { i } - l _ { 0 } ) + d _ { 0 } : d ( l _ { i } , \\lambda _ { d } ) = 0 ,",
"type": "interline_equation",
"image_path": "8c109f14baba8fdb3fa4ca00a590e185e17699666a46300fe54ae3c09558dc5d.jpg"
}
]
}
],
"index": 40,
"virtual_lines": [
{
"bbox": [
171,
624,
438,
639
],
"spans": [],
"index": 40
}
]
},
{
"type": "text",
"bbox": [
106,
643,
506,
732
],
"lines": [
{
"bbox": [
105,
643,
507,
657
],
"spans": [
{
"bbox": [
105,
643,
133,
657
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
133,
644,
183,
655
],
"score": 0.9,
"content": "d _ { 0 } = 2 . 8 \\mathrm { m s }",
"type": "inline_equation"
},
{
"bbox": [
184,
643,
277,
657
],
"score": 1.0,
"content": "is the delay at distance",
"type": "text"
},
{
"bbox": [
278,
644,
324,
655
],
"score": 0.93,
"content": "l _ { 0 } = 1 0 \\mathrm { c m }",
"type": "inline_equation"
},
{
"bbox": [
324,
643,
507,
657
],
"score": 1.0,
"content": ". The delay depends not only on the distance,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
655,
505,
667
],
"spans": [
{
"bbox": [
106,
655,
505,
667
],
"score": 1.0,
"content": "but also on other parameters such as the electrode-to-skin contact, the humidity of the skin, etc. To",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
665,
506,
678
],
"spans": [
{
"bbox": [
105,
665,
506,
678
],
"score": 1.0,
"content": "this end, we evaluate the propagation of the attack with different characteristic delay parameters",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
677,
506,
689
],
"spans": [
{
"bbox": [
106,
677,
192,
689
],
"score": 0.88,
"content": "\\lambda _ { d } \\in \\left[ 0 . 1 , 0 . 5 6 3 \\right] \\mathrm { s / m }",
"type": "inline_equation"
},
{
"bbox": [
192,
677,
218,
689
],
"score": 1.0,
"content": ". With",
"type": "text"
},
{
"bbox": [
218,
677,
256,
688
],
"score": 0.93,
"content": "\\lambda _ { d } = 0 . 1",
"type": "inline_equation"
},
{
"bbox": [
257,
677,
506,
689
],
"score": 1.0,
"content": "we cover the cases where very little delay happens, while the",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
688,
506,
700
],
"spans": [
{
"bbox": [
105,
688,
183,
700
],
"score": 1.0,
"content": "largest considered",
"type": "text"
},
{
"bbox": [
183,
688,
249,
699
],
"score": 0.9,
"content": "\\lambda _ { d } = 0 . 5 6 3 \\mathrm { s / m }",
"type": "inline_equation"
},
{
"bbox": [
249,
688,
365,
700
],
"score": 1.0,
"content": "yields a maximum delay of",
"type": "text"
},
{
"bbox": [
365,
688,
385,
698
],
"score": 0.34,
"content": "0 . 1 \\mathrm { s }",
"type": "inline_equation"
},
{
"bbox": [
385,
688,
506,
700
],
"score": 1.0,
"content": "at the farthest electrode T10,",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
104,
698,
507,
712
],
"spans": [
{
"bbox": [
104,
698,
507,
712
],
"score": 1.0,
"content": "which is in alignment with the observed EEG measurements (Merlet et al., 2013; Sazgar & Young,",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
106,
710,
505,
722
],
"spans": [
{
"bbox": [
106,
710,
185,
722
],
"score": 1.0,
"content": "2019). Similarly to",
"type": "text"
},
{
"bbox": [
185,
710,
199,
721
],
"score": 0.88,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
199,
710,
389,
722
],
"score": 1.0,
"content": ", we showcase also for an intermediate value of",
"type": "text"
},
{
"bbox": [
389,
710,
427,
721
],
"score": 0.91,
"content": "\\lambda _ { d } = 0 . 3",
"type": "inline_equation"
},
{
"bbox": [
427,
710,
505,
722
],
"score": 1.0,
"content": "which corresponds",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
106,
721,
228,
733
],
"spans": [
{
"bbox": [
106,
721,
158,
733
],
"score": 1.0,
"content": "to a delay of",
"type": "text"
},
{
"bbox": [
159,
721,
195,
731
],
"score": 0.27,
"content": "0 . 0 5 3 \\mathrm { m s }",
"type": "inline_equation"
},
{
"bbox": [
196,
721,
228,
733
],
"score": 1.0,
"content": "at T10.",
"type": "text"
}
],
"index": 48
}
],
"index": 44.5,
"bbox_fs": [
104,
643,
507,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
97,
84,
501,
322
],
"lines": [
{
"bbox": [
106,
84,
347,
97
],
"spans": [
{
"bbox": [
106,
84,
347,
97
],
"score": 1.0,
"content": "Algorithm 1: Generation of physiologically plausible UAP.",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
99,
479,
111
],
"spans": [
{
"bbox": [
106,
99,
140,
111
],
"score": 1.0,
"content": "input :",
"type": "text"
},
{
"bbox": [
141,
99,
169,
110
],
"score": 0.88,
"content": "\\mathbf { X } _ { t r a i n }",
"type": "inline_equation"
},
{
"bbox": [
170,
99,
266,
111
],
"score": 1.0,
"content": ", EEG training samples;",
"type": "text"
},
{
"bbox": [
266,
99,
296,
110
],
"score": 0.83,
"content": "\\lambda _ { m } , \\lambda _ { d }",
"type": "inline_equation"
},
{
"bbox": [
296,
99,
427,
111
],
"score": 1.0,
"content": ", spatial propagation parameters;",
"type": "text"
},
{
"bbox": [
427,
100,
434,
110
],
"score": 0.79,
"content": "\\beta",
"type": "inline_equation"
},
{
"bbox": [
435,
99,
479,
111
],
"score": 1.0,
"content": ", weight of",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
139,
109,
491,
122
],
"spans": [
{
"bbox": [
139,
109,
370,
122
],
"score": 1.0,
"content": "derivative loss term; \u000f, maximum perturbation amplitude;",
"type": "text"
},
{
"bbox": [
370,
110,
379,
120
],
"score": 0.78,
"content": "G",
"type": "inline_equation"
},
{
"bbox": [
380,
109,
491,
122
],
"score": 1.0,
"content": ", number of PGD iterations;",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
141,
120,
227,
133
],
"spans": [
{
"bbox": [
141,
121,
149,
131
],
"score": 0.66,
"content": "E",
"type": "inline_equation"
},
{
"bbox": [
150,
120,
227,
133
],
"score": 1.0,
"content": ", number of epochs",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
131,
249,
144
],
"spans": [
{
"bbox": [
106,
131,
249,
144
],
"score": 1.0,
"content": "output :v, adversarial perturbation",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
96,
146,
490,
159
],
"spans": [
{
"bbox": [
96,
146,
106,
159
],
"score": 1.0,
"content": "1",
"type": "text"
},
{
"bbox": [
106,
146,
197,
159
],
"score": 0.75,
"content": "\\mathbf { v } \\mathcal { U } ( - \\epsilon , \\epsilon ) \\in \\mathbb { R } ^ { N _ { s } } ;",
"type": "inline_equation"
},
{
"bbox": [
198,
146,
200,
159
],
"score": 1.0,
"content": ";",
"type": "text"
},
{
"bbox": [
395,
147,
490,
159
],
"score": 1.0,
"content": "// Initialisation",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
97,
157,
182,
170
],
"spans": [
{
"bbox": [
97,
157,
121,
170
],
"score": 1.0,
"content": "2 for",
"type": "text"
},
{
"bbox": [
122,
160,
148,
169
],
"score": 0.83,
"content": "e \\gets 1",
"type": "inline_equation"
},
{
"bbox": [
148,
157,
159,
170
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
160,
160,
168,
168
],
"score": 0.67,
"content": "E",
"type": "inline_equation"
},
{
"bbox": [
168,
157,
182,
170
],
"score": 1.0,
"content": "do",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
98,
167,
187,
183
],
"spans": [
{
"bbox": [
98,
172,
105,
180
],
"score": 1.0,
"content": "3",
"type": "text"
},
{
"bbox": [
120,
167,
153,
183
],
"score": 1.0,
"content": "Shuffle",
"type": "text"
},
{
"bbox": [
153,
169,
182,
181
],
"score": 0.89,
"content": "\\mathbf { X } _ { t r a i n }",
"type": "inline_equation"
},
{
"bbox": [
182,
167,
187,
183
],
"score": 1.0,
"content": ";",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
99,
178,
249,
193
],
"spans": [
{
"bbox": [
99,
183,
104,
190
],
"score": 1.0,
"content": "4",
"type": "text"
},
{
"bbox": [
120,
178,
185,
193
],
"score": 1.0,
"content": "for each batch",
"type": "text"
},
{
"bbox": [
185,
181,
235,
191
],
"score": 0.86,
"content": "\\mathbf { B } \\in \\mathbf { X } _ { t r a i n }",
"type": "inline_equation"
},
{
"bbox": [
235,
178,
249,
193
],
"score": 1.0,
"content": "do",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
99,
189,
173,
206
],
"spans": [
{
"bbox": [
99,
194,
105,
201
],
"score": 1.0,
"content": "5",
"type": "text"
},
{
"bbox": [
137,
191,
167,
204
],
"score": 0.84,
"content": "\\alpha \\frac \\epsilon 2",
"type": "inline_equation"
},
{
"bbox": [
167,
189,
173,
206
],
"score": 1.0,
"content": ";",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
99,
202,
215,
215
],
"spans": [
{
"bbox": [
99,
207,
105,
214
],
"score": 1.0,
"content": "6",
"type": "text"
},
{
"bbox": [
135,
202,
152,
215
],
"score": 1.0,
"content": "for",
"type": "text"
},
{
"bbox": [
152,
204,
179,
214
],
"score": 0.8,
"content": "g \\gets 1",
"type": "inline_equation"
},
{
"bbox": [
180,
202,
191,
215
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
191,
204,
200,
213
],
"score": 0.67,
"content": "G",
"type": "inline_equation"
},
{
"bbox": [
200,
202,
215,
215
],
"score": 1.0,
"content": "do",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
99,
213,
491,
228
],
"spans": [
{
"bbox": [
99,
217,
105,
225
],
"score": 1.0,
"content": "7",
"type": "text"
},
{
"bbox": [
151,
214,
235,
226
],
"score": 0.3,
"content": "\\mathbf { \\bar { V } } H ( \\mathbf { v } , \\lambda _ { m } , \\lambda _ { d } ) ;",
"type": "inline_equation"
},
{
"bbox": [
366,
213,
491,
228
],
"score": 1.0,
"content": "// Spatial propagation",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
99,
222,
491,
238
],
"spans": [
{
"bbox": [
99,
228,
106,
236
],
"score": 1.0,
"content": "8",
"type": "text"
},
{
"bbox": [
150,
222,
243,
237
],
"score": 0.42,
"content": "\\begin{array} { r } { \\dot { \\mathbf { p } } f ( H _ { b p } ( \\mathbf { B } + \\mathbf { V } ) ) } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
243,
224,
248,
237
],
"score": 1.0,
"content": ";",
"type": "text"
},
{
"bbox": [
318,
224,
491,
238
],
"score": 1.0,
"content": "// Model pass with perturbation",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
99,
237,
491,
257
],
"spans": [
{
"bbox": [
99,
244,
106,
252
],
"score": 1.0,
"content": "9",
"type": "text"
},
{
"bbox": [
151,
237,
351,
257
],
"score": 0.37,
"content": "\\begin{array} { r } { \\mathbf { v } \\mathbf { v } - \\alpha \\cdot \\mathrm { s i g n } ( \\nabla _ { \\mathbf { v } } ( l ( \\mathbf { p } , y _ { r e s t } ) - \\frac { \\beta } { \\epsilon } | | \\mathbf { v } ^ { \\prime } | | _ { 1 } ) ) } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
351,
237,
491,
256
],
"score": 1.0,
"content": "; // Update w/derivative",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
96,
253,
490,
269
],
"spans": [
{
"bbox": [
96,
258,
106,
267
],
"score": 1.0,
"content": "10",
"type": "text"
},
{
"bbox": [
150,
253,
216,
269
],
"score": 1.0,
"content": "v ← clip\u000f (v);",
"type": "text"
},
{
"bbox": [
394,
256,
490,
268
],
"score": 1.0,
"content": "// PGD projection",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
96,
267,
490,
287
],
"spans": [
{
"bbox": [
96,
273,
105,
283
],
"score": 1.0,
"content": "11",
"type": "text"
},
{
"bbox": [
153,
268,
231,
284
],
"score": 0.48,
"content": "\\alpha { \\frac { 0 . 1 - { \\frac { \\epsilon } { 2 } } } { G } } \\cdot g + { \\frac { \\epsilon } { 2 } }",
"type": "inline_equation"
},
{
"bbox": [
232,
267,
239,
287
],
"score": 1.0,
"content": ";",
"type": "text"
},
{
"bbox": [
362,
271,
490,
283
],
"score": 1.0,
"content": "// Learning rate update",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
95,
282,
157,
294
],
"spans": [
{
"bbox": [
95,
285,
105,
294
],
"score": 1.0,
"content": "12",
"type": "text"
},
{
"bbox": [
135,
282,
157,
294
],
"score": 1.0,
"content": "end",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
94,
294,
141,
308
],
"spans": [
{
"bbox": [
94,
296,
105,
308
],
"score": 1.0,
"content": "13",
"type": "text"
},
{
"bbox": [
120,
294,
141,
307
],
"score": 1.0,
"content": "end",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
92,
306,
126,
318
],
"spans": [
{
"bbox": [
92,
306,
126,
318
],
"score": 1.0,
"content": "14 end",
"type": "text"
}
],
"index": 18
}
],
"index": 9
},
{
"type": "title",
"bbox": [
107,
345,
203,
357
],
"lines": [
{
"bbox": [
106,
345,
203,
358
],
"spans": [
{
"bbox": [
106,
345,
203,
358
],
"score": 1.0,
"content": "3.3 ATTACK DESIGN",
"type": "text"
}
],
"index": 19
}
],
"index": 19
},
{
"type": "text",
"bbox": [
107,
366,
506,
410
],
"lines": [
{
"bbox": [
106,
366,
506,
378
],
"spans": [
{
"bbox": [
106,
366,
506,
378
],
"score": 1.0,
"content": "We present practical DoS attacks in BCIs that respect domain constraints such as maximum amplitude,",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
377,
506,
389
],
"spans": [
{
"bbox": [
105,
377,
506,
389
],
"score": 1.0,
"content": "spectral distribution, physiological plausibility, and the spatial propagation of the perturbation. To this",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
388,
505,
401
],
"spans": [
{
"bbox": [
105,
388,
505,
401
],
"score": 1.0,
"content": "end, we formulate a general objective function that contains the spatial propagation, the preprocessing",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
400,
284,
411
],
"spans": [
{
"bbox": [
106,
400,
284,
411
],
"score": 1.0,
"content": "step, and the first order derivative loss term:",
"type": "text"
}
],
"index": 23
}
],
"index": 21.5
},
{
"type": "interline_equation",
"bbox": [
163,
416,
447,
440
],
"lines": [
{
"bbox": [
163,
416,
447,
440
],
"spans": [
{
"bbox": [
163,
416,
447,
440
],
"score": 0.9,
"content": "\\mathcal { L } _ { t o t } \\left( \\mathbf { X } , \\mathbf { v } , \\lambda _ { m } , \\lambda _ { d } \\right) = l \\left( H _ { b p } \\left( \\mathbf { X } + H ( \\mathbf { v } , \\lambda _ { m } , \\lambda _ { d } ) \\right) , y _ { r e s t } \\right) - \\frac { \\beta } { \\epsilon } | | \\mathbf { v } ^ { \\prime } | | _ { 1 } ,",
"type": "interline_equation",
"image_path": "fe16bc6e9801b89541c1b4b4d593a5597c8f9ad0f759f089214314544b703deb.jpg"
}
]
}
],
"index": 24,
"virtual_lines": [
{
"bbox": [
163,
416,
447,
440
],
"spans": [],
"index": 24
}
]
},
{
"type": "text",
"bbox": [
107,
445,
505,
468
],
"lines": [
{
"bbox": [
106,
445,
506,
458
],
"spans": [
{
"bbox": [
106,
445,
134,
458
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
445,
156,
457
],
"score": 0.91,
"content": "l ( \\cdot , \\cdot )",
"type": "inline_equation"
},
{
"bbox": [
156,
445,
409,
458
],
"score": 1.0,
"content": "is the negative log-likelihood loss defined in equation 6 and",
"type": "text"
},
{
"bbox": [
410,
446,
441,
457
],
"score": 0.89,
"content": "\\beta { = } 1 \\mathrm { e } { - } 6",
"type": "inline_equation"
},
{
"bbox": [
441,
445,
506,
458
],
"score": 1.0,
"content": "is a scalar that",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
456,
471,
468
],
"spans": [
{
"bbox": [
105,
456,
471,
468
],
"score": 1.0,
"content": "weights the contribution of the derivative loss term. We compare different attack scenarios:",
"type": "text"
}
],
"index": 26
}
],
"index": 25.5
},
{
"type": "text",
"bbox": [
106,
479,
506,
569
],
"lines": [
{
"bbox": [
105,
479,
506,
493
],
"spans": [
{
"bbox": [
105,
479,
506,
493
],
"score": 1.0,
"content": "Instance-based attacks. A perturbation is computed using either FGSM or PGD based on the",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
490,
506,
504
],
"spans": [
{
"bbox": [
105,
490,
309,
504
],
"score": 1.0,
"content": "knowledge of the currently attacked EEG signal",
"type": "text"
},
{
"bbox": [
309,
491,
319,
501
],
"score": 0.27,
"content": "\\mathbf { X }",
"type": "inline_equation"
},
{
"bbox": [
319,
490,
506,
504
],
"score": 1.0,
"content": ". FGSM computes the perturbation as stated",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
502,
504,
514
],
"spans": [
{
"bbox": [
105,
502,
210,
514
],
"score": 1.0,
"content": "in equation 5, where the",
"type": "text"
},
{
"bbox": [
210,
504,
216,
512
],
"score": 0.65,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
217,
502,
466,
514
],
"score": 1.0,
"content": "defines the perturbation amplitude which is varied between",
"type": "text"
},
{
"bbox": [
466,
502,
504,
513
],
"score": 0.82,
"content": "1 { - } 5 0 \\mathrm { m V } .",
"type": "inline_equation"
}
],
"index": 29
},
{
"bbox": [
105,
513,
506,
525
],
"spans": [
{
"bbox": [
105,
513,
351,
525
],
"score": 1.0,
"content": "Alternatively, we compute the perturbation using PGD with",
"type": "text"
},
{
"bbox": [
351,
513,
376,
523
],
"score": 0.88,
"content": "G { = } 1 0",
"type": "inline_equation"
},
{
"bbox": [
377,
513,
506,
525
],
"score": 1.0,
"content": "iterations, where each iteration",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
524,
504,
536
],
"spans": [
{
"bbox": [
106,
524,
416,
536
],
"score": 1.0,
"content": "consists of a gradient-based update of the perturbation and a projection to the",
"type": "text"
},
{
"bbox": [
416,
524,
433,
535
],
"score": 0.9,
"content": "L _ { \\infty }",
"type": "inline_equation"
},
{
"bbox": [
433,
524,
498,
536
],
"score": 1.0,
"content": "ball with radius",
"type": "text"
},
{
"bbox": [
498,
526,
504,
534
],
"score": 0.66,
"content": "\\epsilon",
"type": "inline_equation"
}
],
"index": 31
},
{
"bbox": [
106,
534,
506,
547
],
"spans": [
{
"bbox": [
106,
535,
239,
547
],
"score": 1.0,
"content": "(see equation 7). The update rate",
"type": "text"
},
{
"bbox": [
239,
537,
247,
545
],
"score": 0.8,
"content": "\\alpha",
"type": "inline_equation"
},
{
"bbox": [
248,
535,
319,
547
],
"score": 1.0,
"content": "is initialized with",
"type": "text"
},
{
"bbox": [
320,
534,
335,
547
],
"score": 0.88,
"content": "\\epsilon / 2",
"type": "inline_equation"
},
{
"bbox": [
335,
535,
506,
547
],
"score": 1.0,
"content": "and linearly decreased with each iteration,",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
546,
506,
558
],
"spans": [
{
"bbox": [
105,
546,
209,
558
],
"score": 1.0,
"content": "reaching a final value of",
"type": "text"
},
{
"bbox": [
209,
546,
241,
556
],
"score": 0.78,
"content": "0 . 1 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
241,
546,
506,
558
],
"score": 1.0,
"content": "at iteration 10. The PGD computation is restarted 5 times with",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
556,
507,
570
],
"spans": [
{
"bbox": [
105,
556,
468,
570
],
"score": 1.0,
"content": "different initial perturbations, which are drawn from a uniform distribution within the range",
"type": "text"
},
{
"bbox": [
469,
556,
503,
569
],
"score": 0.92,
"content": "[ - \\epsilon , + \\epsilon ]",
"type": "inline_equation"
},
{
"bbox": [
503,
556,
507,
570
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 34
}
],
"index": 30.5
},
{
"type": "text",
"bbox": [
106,
580,
504,
603
],
"lines": [
{
"bbox": [
106,
580,
505,
593
],
"spans": [
{
"bbox": [
106,
580,
505,
593
],
"score": 1.0,
"content": "Universal attacks. A universal perturbation is computed for all the samples in the training data.",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
106,
591,
270,
604
],
"spans": [
{
"bbox": [
106,
591,
270,
604
],
"score": 1.0,
"content": "We optimize the UAP objective function",
"type": "text"
}
],
"index": 36
}
],
"index": 35.5
},
{
"type": "interline_equation",
"bbox": [
207,
608,
404,
626
],
"lines": [
{
"bbox": [
207,
608,
404,
626
],
"spans": [
{
"bbox": [
207,
608,
404,
626
],
"score": 0.9,
"content": "\\operatorname* { m i n } _ { \\mathbf { v } } E _ { \\mathbf { X } \\sim D } \\mathcal { L } _ { t o t } \\left( \\mathbf { X } , \\mathbf { v } , \\lambda _ { m } , \\lambda _ { d } \\right) \\quad \\mathrm { ~ s . t . ~ } | | \\mathbf { v } | | _ { \\infty } \\leq \\epsilon",
"type": "interline_equation",
"image_path": "abca44dd6ab1e65427156e0254551c8079b96a3d17410266c428125a6fa67a74.jpg"
}
]
}
],
"index": 37,
"virtual_lines": [
{
"bbox": [
207,
608,
404,
626
],
"spans": [],
"index": 37
}
]
},
{
"type": "text",
"bbox": [
106,
631,
505,
676
],
"lines": [
{
"bbox": [
106,
632,
505,
644
],
"spans": [
{
"bbox": [
106,
632,
505,
644
],
"score": 1.0,
"content": "using batched PGD. We pass a batch of 16 samples together with the current perturbation through",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
641,
506,
655
],
"spans": [
{
"bbox": [
105,
641,
506,
655
],
"score": 1.0,
"content": "the preprocessing and classifier, compute the loss function, and update the perturbation based on",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
653,
506,
666
],
"spans": [
{
"bbox": [
105,
653,
329,
666
],
"score": 1.0,
"content": "the negative gradient with consecutive projection to the",
"type": "text"
},
{
"bbox": [
329,
653,
345,
664
],
"score": 0.92,
"content": "L _ { \\infty }",
"type": "inline_equation"
},
{
"bbox": [
345,
653,
411,
666
],
"score": 1.0,
"content": "ball with radius",
"type": "text"
},
{
"bbox": [
411,
655,
417,
663
],
"score": 0.48,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
417,
653,
506,
666
],
"score": 1.0,
"content": ". This step is repeated",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
107,
663,
475,
676
],
"spans": [
{
"bbox": [
107,
664,
131,
675
],
"score": 0.86,
"content": "G { = } 1 0",
"type": "inline_equation"
},
{
"bbox": [
132,
663,
417,
676
],
"score": 1.0,
"content": "times before processing the next batch. Overall, the UAP is learned for",
"type": "text"
},
{
"bbox": [
417,
664,
442,
674
],
"score": 0.88,
"content": "E { = } 1 0",
"type": "inline_equation"
},
{
"bbox": [
443,
663,
475,
676
],
"score": 1.0,
"content": "epochs.",
"type": "text"
}
],
"index": 41
}
],
"index": 39.5
},
{
"type": "text",
"bbox": [
107,
687,
505,
732
],
"lines": [
{
"bbox": [
105,
687,
506,
700
],
"spans": [
{
"bbox": [
105,
687,
506,
700
],
"score": 1.0,
"content": "Propagation model. We distinguish between three use cases of spatial propagation model, where",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
698,
506,
712
],
"spans": [
{
"bbox": [
105,
698,
506,
712
],
"score": 1.0,
"content": "in all cases either an instance-specific attack or a universal attack can be computed: Case 1) Ignore",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
709,
506,
722
],
"spans": [
{
"bbox": [
105,
709,
324,
722
],
"score": 1.0,
"content": "the propagation model: a multi-channel perturbation",
"type": "text"
},
{
"bbox": [
324,
710,
334,
720
],
"score": 0.69,
"content": "\\mathbf { V }",
"type": "inline_equation"
},
{
"bbox": [
335,
709,
506,
722
],
"score": 1.0,
"content": "is computed, which attacks each channel",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
106,
721,
506,
733
],
"spans": [
{
"bbox": [
106,
721,
235,
733
],
"score": 1.0,
"content": "individually, replacing the terms",
"type": "text"
},
{
"bbox": [
235,
721,
292,
732
],
"score": 0.93,
"content": "H ( \\mathbf { v } , \\lambda _ { m } , \\lambda _ { d } )",
"type": "inline_equation"
},
{
"bbox": [
292,
721,
306,
733
],
"score": 1.0,
"content": "by",
"type": "text"
},
{
"bbox": [
306,
721,
316,
731
],
"score": 0.73,
"content": "\\mathbf { V }",
"type": "inline_equation"
},
{
"bbox": [
316,
721,
333,
733
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
334,
721,
344,
731
],
"score": 0.86,
"content": "\\mathbf { v } ^ { \\prime }",
"type": "inline_equation"
},
{
"bbox": [
344,
721,
357,
733
],
"score": 1.0,
"content": "by",
"type": "text"
},
{
"bbox": [
357,
721,
370,
731
],
"score": 0.85,
"content": "\\mathbf { V } ^ { \\prime }",
"type": "inline_equation"
},
{
"bbox": [
370,
721,
506,
733
],
"score": 1.0,
"content": "in equation 15. Case 2) Consider",
"type": "text"
}
],
"index": 45
}
],
"index": 43.5
}
],
"page_idx": 5,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
309,
760
],
"lines": [
{
"bbox": [
302,
750,
310,
761
],
"spans": [
{
"bbox": [
302,
750,
310,
761
],
"score": 1.0,
"content": "6",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
97,
84,
501,
322
],
"lines": [
{
"bbox": [
106,
84,
347,
97
],
"spans": [
{
"bbox": [
106,
84,
347,
97
],
"score": 1.0,
"content": "Algorithm 1: Generation of physiologically plausible UAP.",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
99,
479,
111
],
"spans": [
{
"bbox": [
106,
99,
140,
111
],
"score": 1.0,
"content": "input :",
"type": "text"
},
{
"bbox": [
141,
99,
169,
110
],
"score": 0.88,
"content": "\\mathbf { X } _ { t r a i n }",
"type": "inline_equation"
},
{
"bbox": [
170,
99,
266,
111
],
"score": 1.0,
"content": ", EEG training samples;",
"type": "text"
},
{
"bbox": [
266,
99,
296,
110
],
"score": 0.83,
"content": "\\lambda _ { m } , \\lambda _ { d }",
"type": "inline_equation"
},
{
"bbox": [
296,
99,
427,
111
],
"score": 1.0,
"content": ", spatial propagation parameters;",
"type": "text"
},
{
"bbox": [
427,
100,
434,
110
],
"score": 0.79,
"content": "\\beta",
"type": "inline_equation"
},
{
"bbox": [
435,
99,
479,
111
],
"score": 1.0,
"content": ", weight of",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
139,
109,
491,
122
],
"spans": [
{
"bbox": [
139,
109,
370,
122
],
"score": 1.0,
"content": "derivative loss term; \u000f, maximum perturbation amplitude;",
"type": "text"
},
{
"bbox": [
370,
110,
379,
120
],
"score": 0.78,
"content": "G",
"type": "inline_equation"
},
{
"bbox": [
380,
109,
491,
122
],
"score": 1.0,
"content": ", number of PGD iterations;",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
141,
120,
227,
133
],
"spans": [
{
"bbox": [
141,
121,
149,
131
],
"score": 0.66,
"content": "E",
"type": "inline_equation"
},
{
"bbox": [
150,
120,
227,
133
],
"score": 1.0,
"content": ", number of epochs",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
131,
249,
144
],
"spans": [
{
"bbox": [
106,
131,
249,
144
],
"score": 1.0,
"content": "output :v, adversarial perturbation",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
96,
146,
490,
159
],
"spans": [
{
"bbox": [
96,
146,
106,
159
],
"score": 1.0,
"content": "1",
"type": "text"
},
{
"bbox": [
106,
146,
197,
159
],
"score": 0.75,
"content": "\\mathbf { v } \\mathcal { U } ( - \\epsilon , \\epsilon ) \\in \\mathbb { R } ^ { N _ { s } } ;",
"type": "inline_equation"
},
{
"bbox": [
198,
146,
200,
159
],
"score": 1.0,
"content": ";",
"type": "text"
},
{
"bbox": [
395,
147,
490,
159
],
"score": 1.0,
"content": "// Initialisation",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
97,
157,
182,
170
],
"spans": [
{
"bbox": [
97,
157,
121,
170
],
"score": 1.0,
"content": "2 for",
"type": "text"
},
{
"bbox": [
122,
160,
148,
169
],
"score": 0.83,
"content": "e \\gets 1",
"type": "inline_equation"
},
{
"bbox": [
148,
157,
159,
170
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
160,
160,
168,
168
],
"score": 0.67,
"content": "E",
"type": "inline_equation"
},
{
"bbox": [
168,
157,
182,
170
],
"score": 1.0,
"content": "do",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
98,
167,
187,
183
],
"spans": [
{
"bbox": [
98,
172,
105,
180
],
"score": 1.0,
"content": "3",
"type": "text"
},
{
"bbox": [
120,
167,
153,
183
],
"score": 1.0,
"content": "Shuffle",
"type": "text"
},
{
"bbox": [
153,
169,
182,
181
],
"score": 0.89,
"content": "\\mathbf { X } _ { t r a i n }",
"type": "inline_equation"
},
{
"bbox": [
182,
167,
187,
183
],
"score": 1.0,
"content": ";",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
99,
178,
249,
193
],
"spans": [
{
"bbox": [
99,
183,
104,
190
],
"score": 1.0,
"content": "4",
"type": "text"
},
{
"bbox": [
120,
178,
185,
193
],
"score": 1.0,
"content": "for each batch",
"type": "text"
},
{
"bbox": [
185,
181,
235,
191
],
"score": 0.86,
"content": "\\mathbf { B } \\in \\mathbf { X } _ { t r a i n }",
"type": "inline_equation"
},
{
"bbox": [
235,
178,
249,
193
],
"score": 1.0,
"content": "do",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
99,
189,
173,
206
],
"spans": [
{
"bbox": [
99,
194,
105,
201
],
"score": 1.0,
"content": "5",
"type": "text"
},
{
"bbox": [
137,
191,
167,
204
],
"score": 0.84,
"content": "\\alpha \\frac \\epsilon 2",
"type": "inline_equation"
},
{
"bbox": [
167,
189,
173,
206
],
"score": 1.0,
"content": ";",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
99,
202,
215,
215
],
"spans": [
{
"bbox": [
99,
207,
105,
214
],
"score": 1.0,
"content": "6",
"type": "text"
},
{
"bbox": [
135,
202,
152,
215
],
"score": 1.0,
"content": "for",
"type": "text"
},
{
"bbox": [
152,
204,
179,
214
],
"score": 0.8,
"content": "g \\gets 1",
"type": "inline_equation"
},
{
"bbox": [
180,
202,
191,
215
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
191,
204,
200,
213
],
"score": 0.67,
"content": "G",
"type": "inline_equation"
},
{
"bbox": [
200,
202,
215,
215
],
"score": 1.0,
"content": "do",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
99,
213,
491,
228
],
"spans": [
{
"bbox": [
99,
217,
105,
225
],
"score": 1.0,
"content": "7",
"type": "text"
},
{
"bbox": [
151,
214,
235,
226
],
"score": 0.3,
"content": "\\mathbf { \\bar { V } } H ( \\mathbf { v } , \\lambda _ { m } , \\lambda _ { d } ) ;",
"type": "inline_equation"
},
{
"bbox": [
366,
213,
491,
228
],
"score": 1.0,
"content": "// Spatial propagation",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
99,
222,
491,
238
],
"spans": [
{
"bbox": [
99,
228,
106,
236
],
"score": 1.0,
"content": "8",
"type": "text"
},
{
"bbox": [
150,
222,
243,
237
],
"score": 0.42,
"content": "\\begin{array} { r } { \\dot { \\mathbf { p } } f ( H _ { b p } ( \\mathbf { B } + \\mathbf { V } ) ) } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
243,
224,
248,
237
],
"score": 1.0,
"content": ";",
"type": "text"
},
{
"bbox": [
318,
224,
491,
238
],
"score": 1.0,
"content": "// Model pass with perturbation",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
99,
237,
491,
257
],
"spans": [
{
"bbox": [
99,
244,
106,
252
],
"score": 1.0,
"content": "9",
"type": "text"
},
{
"bbox": [
151,
237,
351,
257
],
"score": 0.37,
"content": "\\begin{array} { r } { \\mathbf { v } \\mathbf { v } - \\alpha \\cdot \\mathrm { s i g n } ( \\nabla _ { \\mathbf { v } } ( l ( \\mathbf { p } , y _ { r e s t } ) - \\frac { \\beta } { \\epsilon } | | \\mathbf { v } ^ { \\prime } | | _ { 1 } ) ) } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
351,
237,
491,
256
],
"score": 1.0,
"content": "; // Update w/derivative",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
96,
253,
490,
269
],
"spans": [
{
"bbox": [
96,
258,
106,
267
],
"score": 1.0,
"content": "10",
"type": "text"
},
{
"bbox": [
150,
253,
216,
269
],
"score": 1.0,
"content": "v ← clip\u000f (v);",
"type": "text"
},
{
"bbox": [
394,
256,
490,
268
],
"score": 1.0,
"content": "// PGD projection",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
96,
267,
490,
287
],
"spans": [
{
"bbox": [
96,
273,
105,
283
],
"score": 1.0,
"content": "11",
"type": "text"
},
{
"bbox": [
153,
268,
231,
284
],
"score": 0.48,
"content": "\\alpha { \\frac { 0 . 1 - { \\frac { \\epsilon } { 2 } } } { G } } \\cdot g + { \\frac { \\epsilon } { 2 } }",
"type": "inline_equation"
},
{
"bbox": [
232,
267,
239,
287
],
"score": 1.0,
"content": ";",
"type": "text"
},
{
"bbox": [
362,
271,
490,
283
],
"score": 1.0,
"content": "// Learning rate update",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
95,
282,
157,
294
],
"spans": [
{
"bbox": [
95,
285,
105,
294
],
"score": 1.0,
"content": "12",
"type": "text"
},
{
"bbox": [
135,
282,
157,
294
],
"score": 1.0,
"content": "end",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
94,
294,
141,
308
],
"spans": [
{
"bbox": [
94,
296,
105,
308
],
"score": 1.0,
"content": "13",
"type": "text"
},
{
"bbox": [
120,
294,
141,
307
],
"score": 1.0,
"content": "end",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
92,
306,
126,
318
],
"spans": [
{
"bbox": [
92,
306,
126,
318
],
"score": 1.0,
"content": "14 end",
"type": "text"
}
],
"index": 18
}
],
"index": 9,
"bbox_fs": [
92,
84,
491,
318
]
},
{
"type": "title",
"bbox": [
107,
345,
203,
357
],
"lines": [
{
"bbox": [
106,
345,
203,
358
],
"spans": [
{
"bbox": [
106,
345,
203,
358
],
"score": 1.0,
"content": "3.3 ATTACK DESIGN",
"type": "text"
}
],
"index": 19
}
],
"index": 19
},
{
"type": "text",
"bbox": [
107,
366,
506,
410
],
"lines": [
{
"bbox": [
106,
366,
506,
378
],
"spans": [
{
"bbox": [
106,
366,
506,
378
],
"score": 1.0,
"content": "We present practical DoS attacks in BCIs that respect domain constraints such as maximum amplitude,",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
377,
506,
389
],
"spans": [
{
"bbox": [
105,
377,
506,
389
],
"score": 1.0,
"content": "spectral distribution, physiological plausibility, and the spatial propagation of the perturbation. To this",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
388,
505,
401
],
"spans": [
{
"bbox": [
105,
388,
505,
401
],
"score": 1.0,
"content": "end, we formulate a general objective function that contains the spatial propagation, the preprocessing",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
400,
284,
411
],
"spans": [
{
"bbox": [
106,
400,
284,
411
],
"score": 1.0,
"content": "step, and the first order derivative loss term:",
"type": "text"
}
],
"index": 23
}
],
"index": 21.5,
"bbox_fs": [
105,
366,
506,
411
]
},
{
"type": "interline_equation",
"bbox": [
163,
416,
447,
440
],
"lines": [
{
"bbox": [
163,
416,
447,
440
],
"spans": [
{
"bbox": [
163,
416,
447,
440
],
"score": 0.9,
"content": "\\mathcal { L } _ { t o t } \\left( \\mathbf { X } , \\mathbf { v } , \\lambda _ { m } , \\lambda _ { d } \\right) = l \\left( H _ { b p } \\left( \\mathbf { X } + H ( \\mathbf { v } , \\lambda _ { m } , \\lambda _ { d } ) \\right) , y _ { r e s t } \\right) - \\frac { \\beta } { \\epsilon } | | \\mathbf { v } ^ { \\prime } | | _ { 1 } ,",
"type": "interline_equation",
"image_path": "fe16bc6e9801b89541c1b4b4d593a5597c8f9ad0f759f089214314544b703deb.jpg"
}
]
}
],
"index": 24,
"virtual_lines": [
{
"bbox": [
163,
416,
447,
440
],
"spans": [],
"index": 24
}
]
},
{
"type": "text",
"bbox": [
107,
445,
505,
468
],
"lines": [
{
"bbox": [
106,
445,
506,
458
],
"spans": [
{
"bbox": [
106,
445,
134,
458
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
445,
156,
457
],
"score": 0.91,
"content": "l ( \\cdot , \\cdot )",
"type": "inline_equation"
},
{
"bbox": [
156,
445,
409,
458
],
"score": 1.0,
"content": "is the negative log-likelihood loss defined in equation 6 and",
"type": "text"
},
{
"bbox": [
410,
446,
441,
457
],
"score": 0.89,
"content": "\\beta { = } 1 \\mathrm { e } { - } 6",
"type": "inline_equation"
},
{
"bbox": [
441,
445,
506,
458
],
"score": 1.0,
"content": "is a scalar that",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
456,
471,
468
],
"spans": [
{
"bbox": [
105,
456,
471,
468
],
"score": 1.0,
"content": "weights the contribution of the derivative loss term. We compare different attack scenarios:",
"type": "text"
}
],
"index": 26
}
],
"index": 25.5,
"bbox_fs": [
105,
445,
506,
468
]
},
{
"type": "text",
"bbox": [
106,
479,
506,
569
],
"lines": [
{
"bbox": [
105,
479,
506,
493
],
"spans": [
{
"bbox": [
105,
479,
506,
493
],
"score": 1.0,
"content": "Instance-based attacks. A perturbation is computed using either FGSM or PGD based on the",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
490,
506,
504
],
"spans": [
{
"bbox": [
105,
490,
309,
504
],
"score": 1.0,
"content": "knowledge of the currently attacked EEG signal",
"type": "text"
},
{
"bbox": [
309,
491,
319,
501
],
"score": 0.27,
"content": "\\mathbf { X }",
"type": "inline_equation"
},
{
"bbox": [
319,
490,
506,
504
],
"score": 1.0,
"content": ". FGSM computes the perturbation as stated",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
502,
504,
514
],
"spans": [
{
"bbox": [
105,
502,
210,
514
],
"score": 1.0,
"content": "in equation 5, where the",
"type": "text"
},
{
"bbox": [
210,
504,
216,
512
],
"score": 0.65,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
217,
502,
466,
514
],
"score": 1.0,
"content": "defines the perturbation amplitude which is varied between",
"type": "text"
},
{
"bbox": [
466,
502,
504,
513
],
"score": 0.82,
"content": "1 { - } 5 0 \\mathrm { m V } .",
"type": "inline_equation"
}
],
"index": 29
},
{
"bbox": [
105,
513,
506,
525
],
"spans": [
{
"bbox": [
105,
513,
351,
525
],
"score": 1.0,
"content": "Alternatively, we compute the perturbation using PGD with",
"type": "text"
},
{
"bbox": [
351,
513,
376,
523
],
"score": 0.88,
"content": "G { = } 1 0",
"type": "inline_equation"
},
{
"bbox": [
377,
513,
506,
525
],
"score": 1.0,
"content": "iterations, where each iteration",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
524,
504,
536
],
"spans": [
{
"bbox": [
106,
524,
416,
536
],
"score": 1.0,
"content": "consists of a gradient-based update of the perturbation and a projection to the",
"type": "text"
},
{
"bbox": [
416,
524,
433,
535
],
"score": 0.9,
"content": "L _ { \\infty }",
"type": "inline_equation"
},
{
"bbox": [
433,
524,
498,
536
],
"score": 1.0,
"content": "ball with radius",
"type": "text"
},
{
"bbox": [
498,
526,
504,
534
],
"score": 0.66,
"content": "\\epsilon",
"type": "inline_equation"
}
],
"index": 31
},
{
"bbox": [
106,
534,
506,
547
],
"spans": [
{
"bbox": [
106,
535,
239,
547
],
"score": 1.0,
"content": "(see equation 7). The update rate",
"type": "text"
},
{
"bbox": [
239,
537,
247,
545
],
"score": 0.8,
"content": "\\alpha",
"type": "inline_equation"
},
{
"bbox": [
248,
535,
319,
547
],
"score": 1.0,
"content": "is initialized with",
"type": "text"
},
{
"bbox": [
320,
534,
335,
547
],
"score": 0.88,
"content": "\\epsilon / 2",
"type": "inline_equation"
},
{
"bbox": [
335,
535,
506,
547
],
"score": 1.0,
"content": "and linearly decreased with each iteration,",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
546,
506,
558
],
"spans": [
{
"bbox": [
105,
546,
209,
558
],
"score": 1.0,
"content": "reaching a final value of",
"type": "text"
},
{
"bbox": [
209,
546,
241,
556
],
"score": 0.78,
"content": "0 . 1 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
241,
546,
506,
558
],
"score": 1.0,
"content": "at iteration 10. The PGD computation is restarted 5 times with",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
556,
507,
570
],
"spans": [
{
"bbox": [
105,
556,
468,
570
],
"score": 1.0,
"content": "different initial perturbations, which are drawn from a uniform distribution within the range",
"type": "text"
},
{
"bbox": [
469,
556,
503,
569
],
"score": 0.92,
"content": "[ - \\epsilon , + \\epsilon ]",
"type": "inline_equation"
},
{
"bbox": [
503,
556,
507,
570
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 34
}
],
"index": 30.5,
"bbox_fs": [
105,
479,
507,
570
]
},
{
"type": "text",
"bbox": [
106,
580,
504,
603
],
"lines": [
{
"bbox": [
106,
580,
505,
593
],
"spans": [
{
"bbox": [
106,
580,
505,
593
],
"score": 1.0,
"content": "Universal attacks. A universal perturbation is computed for all the samples in the training data.",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
106,
591,
270,
604
],
"spans": [
{
"bbox": [
106,
591,
270,
604
],
"score": 1.0,
"content": "We optimize the UAP objective function",
"type": "text"
}
],
"index": 36
}
],
"index": 35.5,
"bbox_fs": [
106,
580,
505,
604
]
},
{
"type": "interline_equation",
"bbox": [
207,
608,
404,
626
],
"lines": [
{
"bbox": [
207,
608,
404,
626
],
"spans": [
{
"bbox": [
207,
608,
404,
626
],
"score": 0.9,
"content": "\\operatorname* { m i n } _ { \\mathbf { v } } E _ { \\mathbf { X } \\sim D } \\mathcal { L } _ { t o t } \\left( \\mathbf { X } , \\mathbf { v } , \\lambda _ { m } , \\lambda _ { d } \\right) \\quad \\mathrm { ~ s . t . ~ } | | \\mathbf { v } | | _ { \\infty } \\leq \\epsilon",
"type": "interline_equation",
"image_path": "abca44dd6ab1e65427156e0254551c8079b96a3d17410266c428125a6fa67a74.jpg"
}
]
}
],
"index": 37,
"virtual_lines": [
{
"bbox": [
207,
608,
404,
626
],
"spans": [],
"index": 37
}
]
},
{
"type": "text",
"bbox": [
106,
631,
505,
676
],
"lines": [
{
"bbox": [
106,
632,
505,
644
],
"spans": [
{
"bbox": [
106,
632,
505,
644
],
"score": 1.0,
"content": "using batched PGD. We pass a batch of 16 samples together with the current perturbation through",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
641,
506,
655
],
"spans": [
{
"bbox": [
105,
641,
506,
655
],
"score": 1.0,
"content": "the preprocessing and classifier, compute the loss function, and update the perturbation based on",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
653,
506,
666
],
"spans": [
{
"bbox": [
105,
653,
329,
666
],
"score": 1.0,
"content": "the negative gradient with consecutive projection to the",
"type": "text"
},
{
"bbox": [
329,
653,
345,
664
],
"score": 0.92,
"content": "L _ { \\infty }",
"type": "inline_equation"
},
{
"bbox": [
345,
653,
411,
666
],
"score": 1.0,
"content": "ball with radius",
"type": "text"
},
{
"bbox": [
411,
655,
417,
663
],
"score": 0.48,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
417,
653,
506,
666
],
"score": 1.0,
"content": ". This step is repeated",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
107,
663,
475,
676
],
"spans": [
{
"bbox": [
107,
664,
131,
675
],
"score": 0.86,
"content": "G { = } 1 0",
"type": "inline_equation"
},
{
"bbox": [
132,
663,
417,
676
],
"score": 1.0,
"content": "times before processing the next batch. Overall, the UAP is learned for",
"type": "text"
},
{
"bbox": [
417,
664,
442,
674
],
"score": 0.88,
"content": "E { = } 1 0",
"type": "inline_equation"
},
{
"bbox": [
443,
663,
475,
676
],
"score": 1.0,
"content": "epochs.",
"type": "text"
}
],
"index": 41
}
],
"index": 39.5,
"bbox_fs": [
105,
632,
506,
676
]
},
{
"type": "text",
"bbox": [
107,
687,
505,
732
],
"lines": [
{
"bbox": [
105,
687,
506,
700
],
"spans": [
{
"bbox": [
105,
687,
506,
700
],
"score": 1.0,
"content": "Propagation model. We distinguish between three use cases of spatial propagation model, where",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
698,
506,
712
],
"spans": [
{
"bbox": [
105,
698,
506,
712
],
"score": 1.0,
"content": "in all cases either an instance-specific attack or a universal attack can be computed: Case 1) Ignore",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
709,
506,
722
],
"spans": [
{
"bbox": [
105,
709,
324,
722
],
"score": 1.0,
"content": "the propagation model: a multi-channel perturbation",
"type": "text"
},
{
"bbox": [
324,
710,
334,
720
],
"score": 0.69,
"content": "\\mathbf { V }",
"type": "inline_equation"
},
{
"bbox": [
335,
709,
506,
722
],
"score": 1.0,
"content": "is computed, which attacks each channel",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
106,
721,
506,
733
],
"spans": [
{
"bbox": [
106,
721,
235,
733
],
"score": 1.0,
"content": "individually, replacing the terms",
"type": "text"
},
{
"bbox": [
235,
721,
292,
732
],
"score": 0.93,
"content": "H ( \\mathbf { v } , \\lambda _ { m } , \\lambda _ { d } )",
"type": "inline_equation"
},
{
"bbox": [
292,
721,
306,
733
],
"score": 1.0,
"content": "by",
"type": "text"
},
{
"bbox": [
306,
721,
316,
731
],
"score": 0.73,
"content": "\\mathbf { V }",
"type": "inline_equation"
},
{
"bbox": [
316,
721,
333,
733
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
334,
721,
344,
731
],
"score": 0.86,
"content": "\\mathbf { v } ^ { \\prime }",
"type": "inline_equation"
},
{
"bbox": [
344,
721,
357,
733
],
"score": 1.0,
"content": "by",
"type": "text"
},
{
"bbox": [
357,
721,
370,
731
],
"score": 0.85,
"content": "\\mathbf { V } ^ { \\prime }",
"type": "inline_equation"
},
{
"bbox": [
370,
721,
506,
733
],
"score": 1.0,
"content": "in equation 15. Case 2) Consider",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
377,
505,
390
],
"spans": [
{
"bbox": [
105,
377,
505,
390
],
"score": 1.0,
"content": "the propagation model: a single-channel perturbation v is computed and tested with a specific",
"type": "text",
"cross_page": true
}
],
"index": 28
},
{
"bbox": [
105,
389,
506,
401
],
"spans": [
{
"bbox": [
105,
389,
210,
401
],
"score": 1.0,
"content": "propagation configuration",
"type": "text",
"cross_page": true
},
{
"bbox": [
210,
389,
225,
400
],
"score": 0.89,
"content": "\\lambda _ { m }",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
225,
389,
243,
401
],
"score": 1.0,
"content": "and",
"type": "text",
"cross_page": true
},
{
"bbox": [
243,
389,
254,
400
],
"score": 0.87,
"content": "\\lambda _ { d }",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
255,
389,
506,
401
],
"score": 1.0,
"content": ". Case 3) Consider a use-case where the attacker does not know",
"type": "text",
"cross_page": true
}
],
"index": 29
},
{
"bbox": [
105,
399,
506,
412
],
"spans": [
{
"bbox": [
105,
399,
380,
412
],
"score": 1.0,
"content": "the spatial propagation model and computes the same perturbation",
"type": "text",
"cross_page": true
},
{
"bbox": [
381,
402,
389,
410
],
"score": 0.32,
"content": "\\mathbf { v }",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
389,
399,
506,
412
],
"score": 1.0,
"content": "for all channels. The actual",
"type": "text",
"cross_page": true
}
],
"index": 30
},
{
"bbox": [
105,
411,
457,
423
],
"spans": [
{
"bbox": [
105,
411,
457,
423
],
"score": 1.0,
"content": "propagation model is applied during testing to model the real-world signal propagation.",
"type": "text",
"cross_page": true
}
],
"index": 31
}
],
"index": 43.5,
"bbox_fs": [
105,
687,
506,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
112,
82,
293,
208
],
"blocks": [
{
"type": "image_body",
"bbox": [
112,
82,
293,
208
],
"group_id": 0,
"lines": [
{
"bbox": [
112,
82,
293,
208
],
"spans": [
{
"bbox": [
112,
82,
293,
208
],
"score": 0.968,
"type": "image",
"image_path": "a1324eec3369ab8d3e852dd7a1f3de592be68aba36ad56c69f815c71827712f0.jpg"
}
]
}
],
"index": 4.5,
"virtual_lines": [
{
"bbox": [
112,
82,
293,
94.6
],
"spans": [],
"index": 0
},
{
"bbox": [
112,
94.6,
293,
107.19999999999999
],
"spans": [],
"index": 1
},
{
"bbox": [
112,
107.19999999999999,
293,
119.79999999999998
],
"spans": [],
"index": 2
},
{
"bbox": [
112,
119.79999999999998,
293,
132.39999999999998
],
"spans": [],
"index": 3
},
{
"bbox": [
112,
132.39999999999998,
293,
144.99999999999997
],
"spans": [],
"index": 4
},
{
"bbox": [
112,
144.99999999999997,
293,
157.59999999999997
],
"spans": [],
"index": 5
},
{
"bbox": [
112,
157.59999999999997,
293,
170.19999999999996
],
"spans": [],
"index": 6
},
{
"bbox": [
112,
170.19999999999996,
293,
182.79999999999995
],
"spans": [],
"index": 7
},
{
"bbox": [
112,
182.79999999999995,
293,
195.39999999999995
],
"spans": [],
"index": 8
},
{
"bbox": [
112,
195.39999999999995,
293,
207.99999999999994
],
"spans": [],
"index": 9
}
]
},
{
"type": "image_caption",
"bbox": [
106,
214,
303,
234
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
214,
304,
225
],
"spans": [
{
"bbox": [
106,
214,
304,
225
],
"score": 1.0,
"content": "Figure 1: Performance of random noise, FGSM, PGD,",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
223,
279,
235
],
"spans": [
{
"bbox": [
106,
223,
279,
235
],
"score": 1.0,
"content": "and UAP with and without derivative loss term.",
"type": "text"
}
],
"index": 11
}
],
"index": 10.5
}
],
"index": 7.5
},
{
"type": "table",
"bbox": [
308,
96,
505,
165
],
"blocks": [
{
"type": "table_body",
"bbox": [
308,
96,
505,
165
],
"group_id": 0,
"lines": [
{
"bbox": [
308,
96,
505,
165
],
"spans": [
{
"bbox": [
308,
96,
505,
165
],
"score": 0.97,
"html": " | n[10-3v2] | l2-norm [mV] | γ[%] |
| ε[mV] | (a) | (b) | (c) | (a) | (b) | (c) | (a) | (b) | （c） |
| 1 | 3.31 | 1.98 | 3.42 | 20.8 | 15.2 | 21.2 | 99.89 | 99.93 | 99.89 |
| 5 | 16.6 | 7.76 | 17.3 | 99.1 | 61.2 | 102 | 97.99 | 99.22 | 97.87 |
| 10 | 32.5 | 12.6 | 34.1 | 191 | 112 | 198 | 93.82 | 97.47 | 93.49 |
| 25 | 74.9 | 27.3 | 79.2 | 461 | 263 | 475 | 79.61 | 90.05 | 78.92 |
| 50 | 125 | 39.0 | 135 | 823 | 462 | 855 | 64.17 | 79.70 | 63.06 |
",
"type": "table",
"image_path": "e61df40a1bdb0c2d0c3c4f3479777943c9d67c63241455dde0304702d2c127b4.jpg"
}
]
}
],
"index": 14,
"virtual_lines": [
{
"bbox": [
308,
96,
505,
109.8
],
"spans": [],
"index": 12
},
{
"bbox": [
308,
109.8,
505,
123.6
],
"spans": [],
"index": 13
},
{
"bbox": [
308,
123.6,
505,
137.4
],
"spans": [],
"index": 14
},
{
"bbox": [
308,
137.4,
505,
151.20000000000002
],
"spans": [],
"index": 15
},
{
"bbox": [
308,
151.20000000000002,
505,
165.00000000000003
],
"spans": [],
"index": 16
}
]
},
{
"type": "table_caption",
"bbox": [
308,
174,
505,
234
],
"group_id": 0,
"lines": [
{
"bbox": [
308,
174,
505,
185
],
"spans": [
{
"bbox": [
308,
174,
505,
185
],
"score": 1.0,
"content": "Table 1: Plausibility metrics for PGD attack (a) without",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
308,
184,
505,
195
],
"spans": [
{
"bbox": [
308,
184,
505,
195
],
"score": 1.0,
"content": "derivative term, (b) with the derivative loss term, and (c)",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
308,
194,
505,
204
],
"spans": [
{
"bbox": [
308,
194,
505,
204
],
"score": 1.0,
"content": "with a Gaussian kernel (Han et al., 2020). The smaller",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
308,
204,
506,
215
],
"spans": [
{
"bbox": [
308,
204,
385,
215
],
"score": 1.0,
"content": "the cross correlation",
"type": "text"
},
{
"bbox": [
386,
206,
392,
214
],
"score": 0.75,
"content": "\\eta",
"type": "inline_equation"
},
{
"bbox": [
393,
204,
492,
215
],
"score": 1.0,
"content": "and the Euclidian distance",
"type": "text"
},
{
"bbox": [
493,
204,
502,
214
],
"score": 0.81,
"content": "\\ell _ { 2 }",
"type": "inline_equation"
},
{
"bbox": [
502,
204,
506,
215
],
"score": 1.0,
"content": "-",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
308,
214,
505,
226
],
"spans": [
{
"bbox": [
308,
214,
461,
226
],
"score": 1.0,
"content": "norm, and the higher the cosine similarity",
"type": "text"
},
{
"bbox": [
461,
216,
468,
225
],
"score": 0.78,
"content": "\\gamma",
"type": "inline_equation"
},
{
"bbox": [
468,
214,
505,
226
],
"score": 1.0,
"content": ", the more",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
307,
224,
412,
234
],
"spans": [
{
"bbox": [
307,
224,
412,
234
],
"score": 1.0,
"content": "natural the generated attack.",
"type": "text"
}
],
"index": 22
}
],
"index": 19.5
}
],
"index": 16.75
},
{
"type": "image",
"bbox": [
109,
247,
504,
325
],
"blocks": [
{
"type": "image_body",
"bbox": [
109,
247,
504,
325
],
"group_id": 1,
"lines": [
{
"bbox": [
109,
247,
504,
325
],
"spans": [
{
"bbox": [
109,
247,
504,
325
],
"score": 0.964,
"type": "image",
"image_path": "0be065a5743b1cee7251d7d88da8df12f3001691d68088f221c3d7a7df59ba26.jpg"
}
]
}
],
"index": 24,
"virtual_lines": [
{
"bbox": [
109,
247,
504,
273.0
],
"spans": [],
"index": 23
},
{
"bbox": [
109,
273.0,
504,
299.0
],
"spans": [],
"index": 24
},
{
"bbox": [
109,
299.0,
504,
325.0
],
"spans": [],
"index": 25
}
]
},
{
"type": "image_caption",
"bbox": [
106,
335,
505,
357
],
"group_id": 1,
"lines": [
{
"bbox": [
106,
335,
505,
347
],
"spans": [
{
"bbox": [
106,
335,
404,
347
],
"score": 1.0,
"content": "Figure 2: A successful attack (a) without and (b) with derivative loss term (PGD,",
"type": "text"
},
{
"bbox": [
404,
336,
437,
345
],
"score": 0.86,
"content": "\\scriptstyle \\epsilon = 1 0 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
438,
335,
505,
347
],
"score": 1.0,
"content": "). The background",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
345,
329,
357
],
"spans": [
{
"bbox": [
105,
345,
329,
357
],
"score": 1.0,
"content": "traces show the original signal before the preprocessing filter.",
"type": "text"
}
],
"index": 27
}
],
"index": 26.5
}
],
"index": 25.25
},
{
"type": "text",
"bbox": [
107,
377,
505,
422
],
"lines": [
{
"bbox": [
105,
377,
505,
390
],
"spans": [
{
"bbox": [
105,
377,
505,
390
],
"score": 1.0,
"content": "the propagation model: a single-channel perturbation v is computed and tested with a specific",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
389,
506,
401
],
"spans": [
{
"bbox": [
105,
389,
210,
401
],
"score": 1.0,
"content": "propagation configuration",
"type": "text"
},
{
"bbox": [
210,
389,
225,
400
],
"score": 0.89,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
225,
389,
243,
401
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
243,
389,
254,
400
],
"score": 0.87,
"content": "\\lambda _ { d }",
"type": "inline_equation"
},
{
"bbox": [
255,
389,
506,
401
],
"score": 1.0,
"content": ". Case 3) Consider a use-case where the attacker does not know",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
399,
506,
412
],
"spans": [
{
"bbox": [
105,
399,
380,
412
],
"score": 1.0,
"content": "the spatial propagation model and computes the same perturbation",
"type": "text"
},
{
"bbox": [
381,
402,
389,
410
],
"score": 0.32,
"content": "\\mathbf { v }",
"type": "inline_equation"
},
{
"bbox": [
389,
399,
506,
412
],
"score": 1.0,
"content": "for all channels. The actual",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
411,
457,
423
],
"spans": [
{
"bbox": [
105,
411,
457,
423
],
"score": 1.0,
"content": "propagation model is applied during testing to model the real-world signal propagation.",
"type": "text"
}
],
"index": 31
}
],
"index": 29.5
},
{
"type": "text",
"bbox": [
107,
434,
505,
489
],
"lines": [
{
"bbox": [
105,
434,
505,
447
],
"spans": [
{
"bbox": [
105,
434,
505,
447
],
"score": 1.0,
"content": "End-to-end algorithm. We illustrate the algorithmic procedure for designing a physiologically",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
445,
505,
457
],
"spans": [
{
"bbox": [
105,
445,
505,
457
],
"score": 1.0,
"content": "plausible UAP in Algorithm 1. Analogously, the proposed methods of derivative loss term and model",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
456,
506,
469
],
"spans": [
{
"bbox": [
105,
456,
332,
469
],
"score": 1.0,
"content": "propagation are applied with PGD. The hyperparameters",
"type": "text"
},
{
"bbox": [
332,
458,
340,
466
],
"score": 0.58,
"content": "\\alpha",
"type": "inline_equation"
},
{
"bbox": [
340,
456,
343,
469
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
344,
457,
351,
468
],
"score": 0.72,
"content": "\\beta",
"type": "inline_equation"
},
{
"bbox": [
352,
456,
506,
469
],
"score": 1.0,
"content": ", the number of PGD iterations and the",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
467,
506,
479
],
"spans": [
{
"bbox": [
105,
467,
506,
479
],
"score": 1.0,
"content": "restarts, the batch size and the number of epochs in UAP are determined based on a cross-validated",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
478,
229,
491
],
"spans": [
{
"bbox": [
105,
478,
229,
491
],
"score": 1.0,
"content": "grid search on the training set.",
"type": "text"
}
],
"index": 36
}
],
"index": 34
},
{
"type": "title",
"bbox": [
108,
506,
272,
518
],
"lines": [
{
"bbox": [
105,
505,
275,
520
],
"spans": [
{
"bbox": [
105,
505,
275,
520
],
"score": 1.0,
"content": "4 EXPERIMENTS AND RESULTS",
"type": "text"
}
],
"index": 37
}
],
"index": 37
},
{
"type": "text",
"bbox": [
107,
530,
505,
564
],
"lines": [
{
"bbox": [
105,
530,
506,
543
],
"spans": [
{
"bbox": [
105,
530,
506,
543
],
"score": 1.0,
"content": "We evaluate our methods on the Physionet EEG Motor Movement/Imagery Dataset (Goldberger et al.,",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
541,
505,
555
],
"spans": [
{
"bbox": [
105,
541,
505,
555
],
"score": 1.0,
"content": "2000; Sch) tackling inter-subject challenges, and generalize to subject-specific inter-session dataset",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
552,
379,
566
],
"spans": [
{
"bbox": [
105,
552,
379,
566
],
"score": 1.0,
"content": "IV-2a of BCI Competition (Brunner et al., 2008) (See Appendix C).",
"type": "text"
}
],
"index": 40
}
],
"index": 39
},
{
"type": "text",
"bbox": [
106,
576,
505,
653
],
"lines": [
{
"bbox": [
106,
577,
506,
589
],
"spans": [
{
"bbox": [
106,
577,
506,
589
],
"score": 1.0,
"content": "Dataset. The Physionet dataset contains valid EEG recordings of 105 subjects (Dose et al., 2018)",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
587,
505,
599
],
"spans": [
{
"bbox": [
106,
587,
505,
599
],
"score": 1.0,
"content": "and is publicly available under Open Data Commons Attribution License v1.0. We use the MI",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
598,
506,
610
],
"spans": [
{
"bbox": [
105,
598,
405,
610
],
"score": 1.0,
"content": "recordings that contain tasks of the imagination of left against right fist for",
"type": "text"
},
{
"bbox": [
405,
599,
417,
609
],
"score": 0.49,
"content": "3 \\mathrm { s }",
"type": "inline_equation"
},
{
"bbox": [
417,
598,
506,
610
],
"score": 1.0,
"content": ". The EEG trials were",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
609,
506,
622
],
"spans": [
{
"bbox": [
105,
609,
165,
622
],
"score": 1.0,
"content": "recorded with",
"type": "text"
},
{
"bbox": [
165,
609,
199,
620
],
"score": 0.91,
"content": "N _ { c h } { = } 6 4",
"type": "inline_equation"
},
{
"bbox": [
200,
609,
284,
622
],
"score": 1.0,
"content": "channels sampled at",
"type": "text"
},
{
"bbox": [
285,
609,
331,
620
],
"score": 0.91,
"content": "F _ { s } { = } 1 6 0 \\mathrm { H z }",
"type": "inline_equation"
},
{
"bbox": [
331,
609,
370,
622
],
"score": 1.0,
"content": ", yielding",
"type": "text"
},
{
"bbox": [
370,
609,
434,
620
],
"score": 0.89,
"content": "N _ { s } { = } 3 { \\cdot } 1 6 0 { = } 4 8 0",
"type": "inline_equation"
},
{
"bbox": [
434,
609,
506,
622
],
"score": 1.0,
"content": "samples per trial.",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
620,
505,
633
],
"spans": [
{
"bbox": [
105,
620,
505,
633
],
"score": 1.0,
"content": "Additional baseline runs provide resting-state data, where the subjects did not perform any tasks",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
104,
631,
507,
644
],
"spans": [
{
"bbox": [
104,
631,
380,
644
],
"score": 1.0,
"content": "while having eyes open. Overall, we get a total of 6615 trials with",
"type": "text"
},
{
"bbox": [
381,
631,
407,
642
],
"score": 0.91,
"content": "N _ { c l } { = } 3",
"type": "inline_equation"
},
{
"bbox": [
408,
631,
507,
644
],
"score": 1.0,
"content": "balanced classes “left”,",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
642,
182,
654
],
"spans": [
{
"bbox": [
105,
642,
182,
654
],
"score": 1.0,
"content": "“right”, and “rest.”",
"type": "text"
}
],
"index": 47
}
],
"index": 44
},
{
"type": "text",
"bbox": [
107,
665,
505,
732
],
"lines": [
{
"bbox": [
106,
665,
505,
678
],
"spans": [
{
"bbox": [
106,
665,
505,
678
],
"score": 1.0,
"content": "Training and validation. We train and validate both the classification models and the generated",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
106,
676,
506,
689
],
"spans": [
{
"bbox": [
106,
676,
506,
689
],
"score": 1.0,
"content": "adversarial examples with a 5-fold cross-validation, splitting the dataset into 84 subjects used for",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
687,
506,
700
],
"spans": [
{
"bbox": [
105,
687,
506,
700
],
"score": 1.0,
"content": "training and 21 subjects used for validation to effectively test the model on inter-subject variability.",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
106,
698,
506,
711
],
"spans": [
{
"bbox": [
106,
698,
506,
711
],
"score": 1.0,
"content": "Similar to Wang et al. (2020), which achieved SoA performance on this dataset, the baseline model is",
"type": "text"
}
],
"index": 51
},
{
"bbox": [
105,
709,
506,
722
],
"spans": [
{
"bbox": [
105,
709,
267,
722
],
"score": 1.0,
"content": "trained for 100 epochs using Adam with",
"type": "text"
},
{
"bbox": [
267,
710,
296,
721
],
"score": 0.86,
"content": "\\beta _ { 1 } { = } 0 . 9",
"type": "inline_equation"
},
{
"bbox": [
297,
709,
300,
722
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
300,
710,
339,
721
],
"score": 0.87,
"content": "\\beta _ { 2 } { = } 0 . 9 9 9",
"type": "inline_equation"
},
{
"bbox": [
339,
709,
506,
722
],
"score": 1.0,
"content": ", and batch size of 16. The learning rate is",
"type": "text"
}
],
"index": 52
},
{
"bbox": [
106,
721,
506,
733
],
"spans": [
{
"bbox": [
106,
721,
471,
733
],
"score": 1.0,
"content": "0.01 and decreased by a factor of 10 at epochs 20 and 50, achieving an average accuracy of",
"type": "text"
},
{
"bbox": [
471,
721,
503,
731
],
"score": 0.88,
"content": "7 4 . 7 8 \\%",
"type": "inline_equation"
},
{
"bbox": [
503,
721,
506,
733
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 53
}
],
"index": 50.5
}
],
"page_idx": 6,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
302,
751,
308,
759
],
"lines": [
{
"bbox": [
302,
750,
309,
762
],
"spans": [
{
"bbox": [
302,
750,
309,
762
],
"score": 1.0,
"content": "7",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
112,
82,
293,
208
],
"blocks": [
{
"type": "image_body",
"bbox": [
112,
82,
293,
208
],
"group_id": 0,
"lines": [
{
"bbox": [
112,
82,
293,
208
],
"spans": [
{
"bbox": [
112,
82,
293,
208
],
"score": 0.968,
"type": "image",
"image_path": "a1324eec3369ab8d3e852dd7a1f3de592be68aba36ad56c69f815c71827712f0.jpg"
}
]
}
],
"index": 4.5,
"virtual_lines": [
{
"bbox": [
112,
82,
293,
94.6
],
"spans": [],
"index": 0
},
{
"bbox": [
112,
94.6,
293,
107.19999999999999
],
"spans": [],
"index": 1
},
{
"bbox": [
112,
107.19999999999999,
293,
119.79999999999998
],
"spans": [],
"index": 2
},
{
"bbox": [
112,
119.79999999999998,
293,
132.39999999999998
],
"spans": [],
"index": 3
},
{
"bbox": [
112,
132.39999999999998,
293,
144.99999999999997
],
"spans": [],
"index": 4
},
{
"bbox": [
112,
144.99999999999997,
293,
157.59999999999997
],
"spans": [],
"index": 5
},
{
"bbox": [
112,
157.59999999999997,
293,
170.19999999999996
],
"spans": [],
"index": 6
},
{
"bbox": [
112,
170.19999999999996,
293,
182.79999999999995
],
"spans": [],
"index": 7
},
{
"bbox": [
112,
182.79999999999995,
293,
195.39999999999995
],
"spans": [],
"index": 8
},
{
"bbox": [
112,
195.39999999999995,
293,
207.99999999999994
],
"spans": [],
"index": 9
}
]
},
{
"type": "image_caption",
"bbox": [
106,
214,
303,
234
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
214,
304,
225
],
"spans": [
{
"bbox": [
106,
214,
304,
225
],
"score": 1.0,
"content": "Figure 1: Performance of random noise, FGSM, PGD,",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
223,
279,
235
],
"spans": [
{
"bbox": [
106,
223,
279,
235
],
"score": 1.0,
"content": "and UAP with and without derivative loss term.",
"type": "text"
}
],
"index": 11
}
],
"index": 10.5
}
],
"index": 7.5
},
{
"type": "table",
"bbox": [
308,
96,
505,
165
],
"blocks": [
{
"type": "table_body",
"bbox": [
308,
96,
505,
165
],
"group_id": 0,
"lines": [
{
"bbox": [
308,
96,
505,
165
],
"spans": [
{
"bbox": [
308,
96,
505,
165
],
"score": 0.97,
"html": " | n[10-3v2] | l2-norm [mV] | γ[%] |
| ε[mV] | (a) | (b) | (c) | (a) | (b) | (c) | (a) | (b) | （c） |
| 1 | 3.31 | 1.98 | 3.42 | 20.8 | 15.2 | 21.2 | 99.89 | 99.93 | 99.89 |
| 5 | 16.6 | 7.76 | 17.3 | 99.1 | 61.2 | 102 | 97.99 | 99.22 | 97.87 |
| 10 | 32.5 | 12.6 | 34.1 | 191 | 112 | 198 | 93.82 | 97.47 | 93.49 |
| 25 | 74.9 | 27.3 | 79.2 | 461 | 263 | 475 | 79.61 | 90.05 | 78.92 |
| 50 | 125 | 39.0 | 135 | 823 | 462 | 855 | 64.17 | 79.70 | 63.06 |
",
"type": "table",
"image_path": "e61df40a1bdb0c2d0c3c4f3479777943c9d67c63241455dde0304702d2c127b4.jpg"
}
]
}
],
"index": 14,
"virtual_lines": [
{
"bbox": [
308,
96,
505,
109.8
],
"spans": [],
"index": 12
},
{
"bbox": [
308,
109.8,
505,
123.6
],
"spans": [],
"index": 13
},
{
"bbox": [
308,
123.6,
505,
137.4
],
"spans": [],
"index": 14
},
{
"bbox": [
308,
137.4,
505,
151.20000000000002
],
"spans": [],
"index": 15
},
{
"bbox": [
308,
151.20000000000002,
505,
165.00000000000003
],
"spans": [],
"index": 16
}
]
},
{
"type": "table_caption",
"bbox": [
308,
174,
505,
234
],
"group_id": 0,
"lines": [
{
"bbox": [
308,
174,
505,
185
],
"spans": [
{
"bbox": [
308,
174,
505,
185
],
"score": 1.0,
"content": "Table 1: Plausibility metrics for PGD attack (a) without",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
308,
184,
505,
195
],
"spans": [
{
"bbox": [
308,
184,
505,
195
],
"score": 1.0,
"content": "derivative term, (b) with the derivative loss term, and (c)",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
308,
194,
505,
204
],
"spans": [
{
"bbox": [
308,
194,
505,
204
],
"score": 1.0,
"content": "with a Gaussian kernel (Han et al., 2020). The smaller",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
308,
204,
506,
215
],
"spans": [
{
"bbox": [
308,
204,
385,
215
],
"score": 1.0,
"content": "the cross correlation",
"type": "text"
},
{
"bbox": [
386,
206,
392,
214
],
"score": 0.75,
"content": "\\eta",
"type": "inline_equation"
},
{
"bbox": [
393,
204,
492,
215
],
"score": 1.0,
"content": "and the Euclidian distance",
"type": "text"
},
{
"bbox": [
493,
204,
502,
214
],
"score": 0.81,
"content": "\\ell _ { 2 }",
"type": "inline_equation"
},
{
"bbox": [
502,
204,
506,
215
],
"score": 1.0,
"content": "-",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
308,
214,
505,
226
],
"spans": [
{
"bbox": [
308,
214,
461,
226
],
"score": 1.0,
"content": "norm, and the higher the cosine similarity",
"type": "text"
},
{
"bbox": [
461,
216,
468,
225
],
"score": 0.78,
"content": "\\gamma",
"type": "inline_equation"
},
{
"bbox": [
468,
214,
505,
226
],
"score": 1.0,
"content": ", the more",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
307,
224,
412,
234
],
"spans": [
{
"bbox": [
307,
224,
412,
234
],
"score": 1.0,
"content": "natural the generated attack.",
"type": "text"
}
],
"index": 22
}
],
"index": 19.5
}
],
"index": 16.75
},
{
"type": "image",
"bbox": [
109,
247,
504,
325
],
"blocks": [
{
"type": "image_body",
"bbox": [
109,
247,
504,
325
],
"group_id": 1,
"lines": [
{
"bbox": [
109,
247,
504,
325
],
"spans": [
{
"bbox": [
109,
247,
504,
325
],
"score": 0.964,
"type": "image",
"image_path": "0be065a5743b1cee7251d7d88da8df12f3001691d68088f221c3d7a7df59ba26.jpg"
}
]
}
],
"index": 24,
"virtual_lines": [
{
"bbox": [
109,
247,
504,
273.0
],
"spans": [],
"index": 23
},
{
"bbox": [
109,
273.0,
504,
299.0
],
"spans": [],
"index": 24
},
{
"bbox": [
109,
299.0,
504,
325.0
],
"spans": [],
"index": 25
}
]
},
{
"type": "image_caption",
"bbox": [
106,
335,
505,
357
],
"group_id": 1,
"lines": [
{
"bbox": [
106,
335,
505,
347
],
"spans": [
{
"bbox": [
106,
335,
404,
347
],
"score": 1.0,
"content": "Figure 2: A successful attack (a) without and (b) with derivative loss term (PGD,",
"type": "text"
},
{
"bbox": [
404,
336,
437,
345
],
"score": 0.86,
"content": "\\scriptstyle \\epsilon = 1 0 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
438,
335,
505,
347
],
"score": 1.0,
"content": "). The background",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
345,
329,
357
],
"spans": [
{
"bbox": [
105,
345,
329,
357
],
"score": 1.0,
"content": "traces show the original signal before the preprocessing filter.",
"type": "text"
}
],
"index": 27
}
],
"index": 26.5
}
],
"index": 25.25
},
{
"type": "text",
"bbox": [
107,
377,
505,
422
],
"lines": [],
"index": 29.5,
"bbox_fs": [
105,
377,
506,
423
],
"lines_deleted": true
},
{
"type": "text",
"bbox": [
107,
434,
505,
489
],
"lines": [
{
"bbox": [
105,
434,
505,
447
],
"spans": [
{
"bbox": [
105,
434,
505,
447
],
"score": 1.0,
"content": "End-to-end algorithm. We illustrate the algorithmic procedure for designing a physiologically",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
445,
505,
457
],
"spans": [
{
"bbox": [
105,
445,
505,
457
],
"score": 1.0,
"content": "plausible UAP in Algorithm 1. Analogously, the proposed methods of derivative loss term and model",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
456,
506,
469
],
"spans": [
{
"bbox": [
105,
456,
332,
469
],
"score": 1.0,
"content": "propagation are applied with PGD. The hyperparameters",
"type": "text"
},
{
"bbox": [
332,
458,
340,
466
],
"score": 0.58,
"content": "\\alpha",
"type": "inline_equation"
},
{
"bbox": [
340,
456,
343,
469
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
344,
457,
351,
468
],
"score": 0.72,
"content": "\\beta",
"type": "inline_equation"
},
{
"bbox": [
352,
456,
506,
469
],
"score": 1.0,
"content": ", the number of PGD iterations and the",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
467,
506,
479
],
"spans": [
{
"bbox": [
105,
467,
506,
479
],
"score": 1.0,
"content": "restarts, the batch size and the number of epochs in UAP are determined based on a cross-validated",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
478,
229,
491
],
"spans": [
{
"bbox": [
105,
478,
229,
491
],
"score": 1.0,
"content": "grid search on the training set.",
"type": "text"
}
],
"index": 36
}
],
"index": 34,
"bbox_fs": [
105,
434,
506,
491
]
},
{
"type": "title",
"bbox": [
108,
506,
272,
518
],
"lines": [
{
"bbox": [
105,
505,
275,
520
],
"spans": [
{
"bbox": [
105,
505,
275,
520
],
"score": 1.0,
"content": "4 EXPERIMENTS AND RESULTS",
"type": "text"
}
],
"index": 37
}
],
"index": 37
},
{
"type": "text",
"bbox": [
107,
530,
505,
564
],
"lines": [
{
"bbox": [
105,
530,
506,
543
],
"spans": [
{
"bbox": [
105,
530,
506,
543
],
"score": 1.0,
"content": "We evaluate our methods on the Physionet EEG Motor Movement/Imagery Dataset (Goldberger et al.,",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
541,
505,
555
],
"spans": [
{
"bbox": [
105,
541,
505,
555
],
"score": 1.0,
"content": "2000; Sch) tackling inter-subject challenges, and generalize to subject-specific inter-session dataset",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
552,
379,
566
],
"spans": [
{
"bbox": [
105,
552,
379,
566
],
"score": 1.0,
"content": "IV-2a of BCI Competition (Brunner et al., 2008) (See Appendix C).",
"type": "text"
}
],
"index": 40
}
],
"index": 39,
"bbox_fs": [
105,
530,
506,
566
]
},
{
"type": "text",
"bbox": [
106,
576,
505,
653
],
"lines": [
{
"bbox": [
106,
577,
506,
589
],
"spans": [
{
"bbox": [
106,
577,
506,
589
],
"score": 1.0,
"content": "Dataset. The Physionet dataset contains valid EEG recordings of 105 subjects (Dose et al., 2018)",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
587,
505,
599
],
"spans": [
{
"bbox": [
106,
587,
505,
599
],
"score": 1.0,
"content": "and is publicly available under Open Data Commons Attribution License v1.0. We use the MI",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
598,
506,
610
],
"spans": [
{
"bbox": [
105,
598,
405,
610
],
"score": 1.0,
"content": "recordings that contain tasks of the imagination of left against right fist for",
"type": "text"
},
{
"bbox": [
405,
599,
417,
609
],
"score": 0.49,
"content": "3 \\mathrm { s }",
"type": "inline_equation"
},
{
"bbox": [
417,
598,
506,
610
],
"score": 1.0,
"content": ". The EEG trials were",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
609,
506,
622
],
"spans": [
{
"bbox": [
105,
609,
165,
622
],
"score": 1.0,
"content": "recorded with",
"type": "text"
},
{
"bbox": [
165,
609,
199,
620
],
"score": 0.91,
"content": "N _ { c h } { = } 6 4",
"type": "inline_equation"
},
{
"bbox": [
200,
609,
284,
622
],
"score": 1.0,
"content": "channels sampled at",
"type": "text"
},
{
"bbox": [
285,
609,
331,
620
],
"score": 0.91,
"content": "F _ { s } { = } 1 6 0 \\mathrm { H z }",
"type": "inline_equation"
},
{
"bbox": [
331,
609,
370,
622
],
"score": 1.0,
"content": ", yielding",
"type": "text"
},
{
"bbox": [
370,
609,
434,
620
],
"score": 0.89,
"content": "N _ { s } { = } 3 { \\cdot } 1 6 0 { = } 4 8 0",
"type": "inline_equation"
},
{
"bbox": [
434,
609,
506,
622
],
"score": 1.0,
"content": "samples per trial.",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
620,
505,
633
],
"spans": [
{
"bbox": [
105,
620,
505,
633
],
"score": 1.0,
"content": "Additional baseline runs provide resting-state data, where the subjects did not perform any tasks",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
104,
631,
507,
644
],
"spans": [
{
"bbox": [
104,
631,
380,
644
],
"score": 1.0,
"content": "while having eyes open. Overall, we get a total of 6615 trials with",
"type": "text"
},
{
"bbox": [
381,
631,
407,
642
],
"score": 0.91,
"content": "N _ { c l } { = } 3",
"type": "inline_equation"
},
{
"bbox": [
408,
631,
507,
644
],
"score": 1.0,
"content": "balanced classes “left”,",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
642,
182,
654
],
"spans": [
{
"bbox": [
105,
642,
182,
654
],
"score": 1.0,
"content": "“right”, and “rest.”",
"type": "text"
}
],
"index": 47
}
],
"index": 44,
"bbox_fs": [
104,
577,
507,
654
]
},
{
"type": "text",
"bbox": [
107,
665,
505,
732
],
"lines": [
{
"bbox": [
106,
665,
505,
678
],
"spans": [
{
"bbox": [
106,
665,
505,
678
],
"score": 1.0,
"content": "Training and validation. We train and validate both the classification models and the generated",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
106,
676,
506,
689
],
"spans": [
{
"bbox": [
106,
676,
506,
689
],
"score": 1.0,
"content": "adversarial examples with a 5-fold cross-validation, splitting the dataset into 84 subjects used for",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
687,
506,
700
],
"spans": [
{
"bbox": [
105,
687,
506,
700
],
"score": 1.0,
"content": "training and 21 subjects used for validation to effectively test the model on inter-subject variability.",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
106,
698,
506,
711
],
"spans": [
{
"bbox": [
106,
698,
506,
711
],
"score": 1.0,
"content": "Similar to Wang et al. (2020), which achieved SoA performance on this dataset, the baseline model is",
"type": "text"
}
],
"index": 51
},
{
"bbox": [
105,
709,
506,
722
],
"spans": [
{
"bbox": [
105,
709,
267,
722
],
"score": 1.0,
"content": "trained for 100 epochs using Adam with",
"type": "text"
},
{
"bbox": [
267,
710,
296,
721
],
"score": 0.86,
"content": "\\beta _ { 1 } { = } 0 . 9",
"type": "inline_equation"
},
{
"bbox": [
297,
709,
300,
722
],
"score": 1.0,
"content": ",",
"type": "text"
},
{
"bbox": [
300,
710,
339,
721
],
"score": 0.87,
"content": "\\beta _ { 2 } { = } 0 . 9 9 9",
"type": "inline_equation"
},
{
"bbox": [
339,
709,
506,
722
],
"score": 1.0,
"content": ", and batch size of 16. The learning rate is",
"type": "text"
}
],
"index": 52
},
{
"bbox": [
106,
721,
506,
733
],
"spans": [
{
"bbox": [
106,
721,
471,
733
],
"score": 1.0,
"content": "0.01 and decreased by a factor of 10 at epochs 20 and 50, achieving an average accuracy of",
"type": "text"
},
{
"bbox": [
471,
721,
503,
731
],
"score": 0.88,
"content": "7 4 . 7 8 \\%",
"type": "inline_equation"
},
{
"bbox": [
503,
721,
506,
733
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 53
}
],
"index": 50.5,
"bbox_fs": [
105,
665,
506,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
108,
81,
502,
209
],
"blocks": [
{
"type": "image_body",
"bbox": [
108,
81,
502,
209
],
"group_id": 0,
"lines": [
{
"bbox": [
108,
81,
502,
209
],
"spans": [
{
"bbox": [
108,
81,
502,
209
],
"score": 0.961,
"type": "image",
"image_path": "61e50ae1624936e4724c0a3920f25f9cef6e945f75c78ede3ed08a1fb5c4512d.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
108,
81,
502,
123.66666666666666
],
"spans": [],
"index": 0
},
{
"bbox": [
108,
123.66666666666666,
502,
166.33333333333331
],
"spans": [],
"index": 1
},
{
"bbox": [
108,
166.33333333333331,
502,
208.99999999999997
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
104,
219,
506,
241
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
218,
506,
231
],
"spans": [
{
"bbox": [
106,
218,
388,
231
],
"score": 1.0,
"content": "Figure 3: A successful attack with derivative loss term and spatial constraints",
"type": "text"
},
{
"bbox": [
389,
220,
419,
230
],
"score": 0.93,
"content": "\\lambda _ { m } = 1",
"type": "inline_equation"
},
{
"bbox": [
419,
218,
435,
231
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
435,
220,
479,
230
],
"score": 0.89,
"content": "\\lambda _ { d } = 0 . 5 6 3",
"type": "inline_equation"
},
{
"bbox": [
480,
218,
506,
231
],
"score": 1.0,
"content": "(PGD,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
230,
428,
241
],
"spans": [
{
"bbox": [
106,
230,
139,
240
],
"score": 0.86,
"content": "\\scriptstyle \\epsilon = 5 0 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
140,
230,
428,
241
],
"score": 1.0,
"content": "). The background traces show the original signal before the preprocessing filter.",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5
}
],
"index": 2.25
},
{
"type": "image",
"bbox": [
110,
256,
379,
500
],
"blocks": [
{
"type": "image_body",
"bbox": [
110,
256,
379,
500
],
"group_id": 1,
"lines": [
{
"bbox": [
110,
256,
379,
500
],
"spans": [
{
"bbox": [
110,
256,
379,
500
],
"score": 0.967,
"type": "image",
"image_path": "1523c7615753b90b2c8e0e94fb41fcac116f3a3069037a001618243cb754a843.jpg"
}
]
}
],
"index": 13.5,
"virtual_lines": [
{
"bbox": [
110,
256,
379,
269.55555555555554
],
"spans": [],
"index": 5
},
{
"bbox": [
110,
269.55555555555554,
379,
283.1111111111111
],
"spans": [],
"index": 6
},
{
"bbox": [
110,
283.1111111111111,
379,
296.66666666666663
],
"spans": [],
"index": 7
},
{
"bbox": [
110,
296.66666666666663,
379,
310.2222222222222
],
"spans": [],
"index": 8
},
{
"bbox": [
110,
310.2222222222222,
379,
323.7777777777777
],
"spans": [],
"index": 9
},
{
"bbox": [
110,
323.7777777777777,
379,
337.33333333333326
],
"spans": [],
"index": 10
},
{
"bbox": [
110,
337.33333333333326,
379,
350.8888888888888
],
"spans": [],
"index": 11
},
{
"bbox": [
110,
350.8888888888888,
379,
364.44444444444434
],
"spans": [],
"index": 12
},
{
"bbox": [
110,
364.44444444444434,
379,
377.9999999999999
],
"spans": [],
"index": 13
},
{
"bbox": [
110,
377.9999999999999,
379,
391.55555555555543
],
"spans": [],
"index": 14
},
{
"bbox": [
110,
391.55555555555543,
379,
405.111111111111
],
"spans": [],
"index": 15
},
{
"bbox": [
110,
405.111111111111,
379,
418.6666666666665
],
"spans": [],
"index": 16
},
{
"bbox": [
110,
418.6666666666665,
379,
432.22222222222206
],
"spans": [],
"index": 17
},
{
"bbox": [
110,
432.22222222222206,
379,
445.7777777777776
],
"spans": [],
"index": 18
},
{
"bbox": [
110,
445.7777777777776,
379,
459.33333333333314
],
"spans": [],
"index": 19
},
{
"bbox": [
110,
459.33333333333314,
379,
472.8888888888887
],
"spans": [],
"index": 20
},
{
"bbox": [
110,
472.8888888888887,
379,
486.44444444444423
],
"spans": [],
"index": 21
},
{
"bbox": [
110,
486.44444444444423,
379,
499.9999999999998
],
"spans": [],
"index": 22
}
]
},
{
"type": "image_caption",
"bbox": [
106,
506,
385,
527
],
"group_id": 1,
"lines": [
{
"bbox": [
106,
506,
387,
518
],
"spans": [
{
"bbox": [
106,
506,
387,
518
],
"score": 1.0,
"content": "Figure 4: ASR of PGD and UAP in Case 2), i.e., computed with head model",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
516,
362,
528
],
"spans": [
{
"bbox": [
105,
516,
362,
528
],
"score": 1.0,
"content": "(w/HM), and in Case 3), i.e., computed without head model (w/oHM).",
"type": "text"
}
],
"index": 24
}
],
"index": 23.5
}
],
"index": 18.5
},
{
"type": "image",
"bbox": [
395,
258,
500,
478
],
"blocks": [
{
"type": "image_body",
"bbox": [
395,
258,
500,
478
],
"group_id": 2,
"lines": [
{
"bbox": [
395,
258,
500,
478
],
"spans": [
{
"bbox": [
395,
258,
500,
478
],
"score": 0.954,
"type": "image",
"image_path": "2c2e639cc82a57badda6f507c714534618e64bf003ecf1957433ac239f2e7d18.jpg"
}
]
}
],
"index": 25,
"virtual_lines": [
{
"bbox": [
395,
258,
500,
478
],
"spans": [],
"index": 25
}
]
},
{
"type": "image_caption",
"bbox": [
391,
486,
505,
527
],
"group_id": 2,
"lines": [
{
"bbox": [
391,
486,
505,
497
],
"spans": [
{
"bbox": [
391,
486,
505,
497
],
"score": 1.0,
"content": "Figure 5: ASR with the",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
390,
495,
506,
508
],
"spans": [
{
"bbox": [
390,
495,
506,
508
],
"score": 1.0,
"content": "PGD attack propagating from",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
391,
506,
505,
516
],
"spans": [
{
"bbox": [
391,
506,
505,
516
],
"score": 1.0,
"content": "different EEG channels with",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
390,
515,
503,
528
],
"spans": [
{
"bbox": [
390,
515,
411,
528
],
"score": 1.0,
"content": "fixed",
"type": "text"
},
{
"bbox": [
412,
516,
439,
526
],
"score": 0.89,
"content": "\\lambda _ { d } { = } 0 . 3",
"type": "inline_equation"
},
{
"bbox": [
439,
515,
486,
528
],
"score": 1.0,
"content": "and variable",
"type": "text"
},
{
"bbox": [
486,
516,
499,
526
],
"score": 0.88,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
500,
515,
503,
528
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 29
}
],
"index": 27.5
}
],
"index": 26.25
},
{
"type": "text",
"bbox": [
107,
551,
505,
596
],
"lines": [
{
"bbox": [
105,
551,
506,
564
],
"spans": [
{
"bbox": [
105,
551,
204,
564
],
"score": 1.0,
"content": "An FFT band-pass filter",
"type": "text"
},
{
"bbox": [
204,
552,
219,
563
],
"score": 0.9,
"content": "h _ { b p }",
"type": "inline_equation"
},
{
"bbox": [
219,
551,
341,
564
],
"score": 1.0,
"content": "with a customary passband of",
"type": "text"
},
{
"bbox": [
341,
551,
383,
562
],
"score": 0.8,
"content": "0 . 1 { - } 4 0 \\mathrm { H z }",
"type": "inline_equation"
},
{
"bbox": [
384,
551,
506,
564
],
"score": 1.0,
"content": "(Lawhern et al., 2018) is used",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
563,
505,
574
],
"spans": [
{
"bbox": [
105,
563,
505,
574
],
"score": 1.0,
"content": "as preprocessing step in both baseline and attack experiments.To determine the ASR, we compute the",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
574,
505,
585
],
"spans": [
{
"bbox": [
106,
574,
505,
585
],
"score": 1.0,
"content": "ratio between the successfully fooled trials, i.e., trials now classified as “rest”, and the total number",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
584,
489,
596
],
"spans": [
{
"bbox": [
105,
584,
489,
596
],
"score": 1.0,
"content": "of attacked trials, where we only consider the ones initially correctly classified as “left”/“right”.",
"type": "text"
}
],
"index": 33
}
],
"index": 31.5
},
{
"type": "text",
"bbox": [
107,
610,
505,
732
],
"lines": [
{
"bbox": [
105,
609,
506,
624
],
"spans": [
{
"bbox": [
105,
609,
506,
624
],
"score": 1.0,
"content": "Physiologically plausible attacks. We first analyze the instance-based attacks without considering",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
622,
505,
634
],
"spans": [
{
"bbox": [
106,
622,
505,
634
],
"score": 1.0,
"content": "the propagation model (Case 1), depicted in Figure 1. We compare our methods against random",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
632,
506,
646
],
"spans": [
{
"bbox": [
105,
632,
191,
646
],
"score": 1.0,
"content": "noise with amplitude",
"type": "text"
},
{
"bbox": [
191,
635,
198,
643
],
"score": 0.43,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
198,
632,
506,
646
],
"score": 1.0,
"content": "as in (Zhang & Wu, 2019), FGSM that is the same as in (Zhang & Wu, 2019)",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
644,
505,
657
],
"spans": [
{
"bbox": [
106,
644,
505,
657
],
"score": 1.0,
"content": "with targeted scenario, and a UAP designed specifically for EEG (Liu et al., 2021). For both FGSM",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
654,
506,
668
],
"spans": [
{
"bbox": [
105,
654,
390,
668
],
"score": 1.0,
"content": "and PGD, the ASR increases together with the maximum amplitude",
"type": "text"
},
{
"bbox": [
390,
657,
396,
665
],
"score": 0.65,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
397,
654,
506,
668
],
"score": 1.0,
"content": "of the perturbation. They",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
665,
505,
678
],
"spans": [
{
"bbox": [
106,
665,
505,
678
],
"score": 1.0,
"content": "always outperform the random noise, with PGD performing slightly better than FGSM. They reach",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
677,
505,
689
],
"spans": [
{
"bbox": [
106,
677,
194,
689
],
"score": 1.0,
"content": "the maximum ASR of",
"type": "text"
},
{
"bbox": [
195,
677,
226,
687
],
"score": 0.88,
"content": "9 9 . 9 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
226,
677,
247,
689
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
247,
677,
273,
687
],
"score": 0.59,
"content": "1 0 \\mathrm { m V } .",
"type": "inline_equation"
},
{
"bbox": [
273,
677,
473,
689
],
"score": 1.0,
"content": "The post-attack classification accuracy drops from",
"type": "text"
},
{
"bbox": [
473,
677,
505,
687
],
"score": 0.86,
"content": "7 4 . 7 8 \\%",
"type": "inline_equation"
}
],
"index": 40
},
{
"bbox": [
105,
687,
505,
700
],
"spans": [
{
"bbox": [
105,
687,
116,
700
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
117,
688,
136,
698
],
"score": 0.86,
"content": "48 \\%",
"type": "inline_equation"
},
{
"bbox": [
136,
687,
259,
700
],
"score": 1.0,
"content": "for a perturbation amplitude of",
"type": "text"
},
{
"bbox": [
260,
688,
283,
698
],
"score": 0.64,
"content": "2 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
283,
687,
310,
700
],
"score": 1.0,
"content": "and to",
"type": "text"
},
{
"bbox": [
310,
688,
329,
698
],
"score": 0.87,
"content": "33 \\%",
"type": "inline_equation"
},
{
"bbox": [
330,
687,
344,
700
],
"score": 1.0,
"content": "for",
"type": "text"
},
{
"bbox": [
345,
688,
372,
698
],
"score": 0.75,
"content": "1 0 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
372,
687,
505,
700
],
"score": 1.0,
"content": "and higher amplitudes. Figure 2a",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
698,
507,
712
],
"spans": [
{
"bbox": [
105,
698,
507,
712
],
"score": 1.0,
"content": "shows the signals of a successful attack using PGD. The adversarial perturbation has a square-",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "wave form which negatively affects the natural shape of the EEG signal. By adding the proposed",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
721,
505,
733
],
"spans": [
{
"bbox": [
106,
721,
505,
733
],
"score": 1.0,
"content": "derivative term, the square-wave artifacts are significantly reduced (2b), making the perturbation more",
"type": "text"
}
],
"index": 44
}
],
"index": 39
}
],
"page_idx": 7,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
302,
752,
308,
760
],
"lines": [
{
"bbox": [
301,
750,
310,
762
],
"spans": [
{
"bbox": [
301,
750,
310,
762
],
"score": 1.0,
"content": "",
"type": "text",
"height": 12,
"width": 9
}
]
}
]
},
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
309,
39
],
"spans": [
{
"bbox": [
106,
25,
309,
39
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
108,
81,
502,
209
],
"blocks": [
{
"type": "image_body",
"bbox": [
108,
81,
502,
209
],
"group_id": 0,
"lines": [
{
"bbox": [
108,
81,
502,
209
],
"spans": [
{
"bbox": [
108,
81,
502,
209
],
"score": 0.961,
"type": "image",
"image_path": "61e50ae1624936e4724c0a3920f25f9cef6e945f75c78ede3ed08a1fb5c4512d.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
108,
81,
502,
123.66666666666666
],
"spans": [],
"index": 0
},
{
"bbox": [
108,
123.66666666666666,
502,
166.33333333333331
],
"spans": [],
"index": 1
},
{
"bbox": [
108,
166.33333333333331,
502,
208.99999999999997
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
104,
219,
506,
241
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
218,
506,
231
],
"spans": [
{
"bbox": [
106,
218,
388,
231
],
"score": 1.0,
"content": "Figure 3: A successful attack with derivative loss term and spatial constraints",
"type": "text"
},
{
"bbox": [
389,
220,
419,
230
],
"score": 0.93,
"content": "\\lambda _ { m } = 1",
"type": "inline_equation"
},
{
"bbox": [
419,
218,
435,
231
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
435,
220,
479,
230
],
"score": 0.89,
"content": "\\lambda _ { d } = 0 . 5 6 3",
"type": "inline_equation"
},
{
"bbox": [
480,
218,
506,
231
],
"score": 1.0,
"content": "(PGD,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
230,
428,
241
],
"spans": [
{
"bbox": [
106,
230,
139,
240
],
"score": 0.86,
"content": "\\scriptstyle \\epsilon = 5 0 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
140,
230,
428,
241
],
"score": 1.0,
"content": "). The background traces show the original signal before the preprocessing filter.",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5
}
],
"index": 2.25
},
{
"type": "image",
"bbox": [
110,
256,
379,
500
],
"blocks": [
{
"type": "image_body",
"bbox": [
110,
256,
379,
500
],
"group_id": 1,
"lines": [
{
"bbox": [
110,
256,
379,
500
],
"spans": [
{
"bbox": [
110,
256,
379,
500
],
"score": 0.967,
"type": "image",
"image_path": "1523c7615753b90b2c8e0e94fb41fcac116f3a3069037a001618243cb754a843.jpg"
}
]
}
],
"index": 13.5,
"virtual_lines": [
{
"bbox": [
110,
256,
379,
269.55555555555554
],
"spans": [],
"index": 5
},
{
"bbox": [
110,
269.55555555555554,
379,
283.1111111111111
],
"spans": [],
"index": 6
},
{
"bbox": [
110,
283.1111111111111,
379,
296.66666666666663
],
"spans": [],
"index": 7
},
{
"bbox": [
110,
296.66666666666663,
379,
310.2222222222222
],
"spans": [],
"index": 8
},
{
"bbox": [
110,
310.2222222222222,
379,
323.7777777777777
],
"spans": [],
"index": 9
},
{
"bbox": [
110,
323.7777777777777,
379,
337.33333333333326
],
"spans": [],
"index": 10
},
{
"bbox": [
110,
337.33333333333326,
379,
350.8888888888888
],
"spans": [],
"index": 11
},
{
"bbox": [
110,
350.8888888888888,
379,
364.44444444444434
],
"spans": [],
"index": 12
},
{
"bbox": [
110,
364.44444444444434,
379,
377.9999999999999
],
"spans": [],
"index": 13
},
{
"bbox": [
110,
377.9999999999999,
379,
391.55555555555543
],
"spans": [],
"index": 14
},
{
"bbox": [
110,
391.55555555555543,
379,
405.111111111111
],
"spans": [],
"index": 15
},
{
"bbox": [
110,
405.111111111111,
379,
418.6666666666665
],
"spans": [],
"index": 16
},
{
"bbox": [
110,
418.6666666666665,
379,
432.22222222222206
],
"spans": [],
"index": 17
},
{
"bbox": [
110,
432.22222222222206,
379,
445.7777777777776
],
"spans": [],
"index": 18
},
{
"bbox": [
110,
445.7777777777776,
379,
459.33333333333314
],
"spans": [],
"index": 19
},
{
"bbox": [
110,
459.33333333333314,
379,
472.8888888888887
],
"spans": [],
"index": 20
},
{
"bbox": [
110,
472.8888888888887,
379,
486.44444444444423
],
"spans": [],
"index": 21
},
{
"bbox": [
110,
486.44444444444423,
379,
499.9999999999998
],
"spans": [],
"index": 22
}
]
},
{
"type": "image_caption",
"bbox": [
106,
506,
385,
527
],
"group_id": 1,
"lines": [
{
"bbox": [
106,
506,
387,
518
],
"spans": [
{
"bbox": [
106,
506,
387,
518
],
"score": 1.0,
"content": "Figure 4: ASR of PGD and UAP in Case 2), i.e., computed with head model",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
516,
362,
528
],
"spans": [
{
"bbox": [
105,
516,
362,
528
],
"score": 1.0,
"content": "(w/HM), and in Case 3), i.e., computed without head model (w/oHM).",
"type": "text"
}
],
"index": 24
}
],
"index": 23.5
}
],
"index": 18.5
},
{
"type": "image",
"bbox": [
395,
258,
500,
478
],
"blocks": [
{
"type": "image_body",
"bbox": [
395,
258,
500,
478
],
"group_id": 2,
"lines": [
{
"bbox": [
395,
258,
500,
478
],
"spans": [
{
"bbox": [
395,
258,
500,
478
],
"score": 0.954,
"type": "image",
"image_path": "2c2e639cc82a57badda6f507c714534618e64bf003ecf1957433ac239f2e7d18.jpg"
}
]
}
],
"index": 25,
"virtual_lines": [
{
"bbox": [
395,
258,
500,
478
],
"spans": [],
"index": 25
}
]
},
{
"type": "image_caption",
"bbox": [
391,
486,
505,
527
],
"group_id": 2,
"lines": [
{
"bbox": [
391,
486,
505,
497
],
"spans": [
{
"bbox": [
391,
486,
505,
497
],
"score": 1.0,
"content": "Figure 5: ASR with the",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
390,
495,
506,
508
],
"spans": [
{
"bbox": [
390,
495,
506,
508
],
"score": 1.0,
"content": "PGD attack propagating from",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
391,
506,
505,
516
],
"spans": [
{
"bbox": [
391,
506,
505,
516
],
"score": 1.0,
"content": "different EEG channels with",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
390,
515,
503,
528
],
"spans": [
{
"bbox": [
390,
515,
411,
528
],
"score": 1.0,
"content": "fixed",
"type": "text"
},
{
"bbox": [
412,
516,
439,
526
],
"score": 0.89,
"content": "\\lambda _ { d } { = } 0 . 3",
"type": "inline_equation"
},
{
"bbox": [
439,
515,
486,
528
],
"score": 1.0,
"content": "and variable",
"type": "text"
},
{
"bbox": [
486,
516,
499,
526
],
"score": 0.88,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
500,
515,
503,
528
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 29
}
],
"index": 27.5
}
],
"index": 26.25
},
{
"type": "text",
"bbox": [
107,
551,
505,
596
],
"lines": [
{
"bbox": [
105,
551,
506,
564
],
"spans": [
{
"bbox": [
105,
551,
204,
564
],
"score": 1.0,
"content": "An FFT band-pass filter",
"type": "text"
},
{
"bbox": [
204,
552,
219,
563
],
"score": 0.9,
"content": "h _ { b p }",
"type": "inline_equation"
},
{
"bbox": [
219,
551,
341,
564
],
"score": 1.0,
"content": "with a customary passband of",
"type": "text"
},
{
"bbox": [
341,
551,
383,
562
],
"score": 0.8,
"content": "0 . 1 { - } 4 0 \\mathrm { H z }",
"type": "inline_equation"
},
{
"bbox": [
384,
551,
506,
564
],
"score": 1.0,
"content": "(Lawhern et al., 2018) is used",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
563,
505,
574
],
"spans": [
{
"bbox": [
105,
563,
505,
574
],
"score": 1.0,
"content": "as preprocessing step in both baseline and attack experiments.To determine the ASR, we compute the",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
574,
505,
585
],
"spans": [
{
"bbox": [
106,
574,
505,
585
],
"score": 1.0,
"content": "ratio between the successfully fooled trials, i.e., trials now classified as “rest”, and the total number",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
584,
489,
596
],
"spans": [
{
"bbox": [
105,
584,
489,
596
],
"score": 1.0,
"content": "of attacked trials, where we only consider the ones initially correctly classified as “left”/“right”.",
"type": "text"
}
],
"index": 33
}
],
"index": 31.5,
"bbox_fs": [
105,
551,
506,
596
]
},
{
"type": "text",
"bbox": [
107,
610,
505,
732
],
"lines": [
{
"bbox": [
105,
609,
506,
624
],
"spans": [
{
"bbox": [
105,
609,
506,
624
],
"score": 1.0,
"content": "Physiologically plausible attacks. We first analyze the instance-based attacks without considering",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
622,
505,
634
],
"spans": [
{
"bbox": [
106,
622,
505,
634
],
"score": 1.0,
"content": "the propagation model (Case 1), depicted in Figure 1. We compare our methods against random",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
632,
506,
646
],
"spans": [
{
"bbox": [
105,
632,
191,
646
],
"score": 1.0,
"content": "noise with amplitude",
"type": "text"
},
{
"bbox": [
191,
635,
198,
643
],
"score": 0.43,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
198,
632,
506,
646
],
"score": 1.0,
"content": "as in (Zhang & Wu, 2019), FGSM that is the same as in (Zhang & Wu, 2019)",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
644,
505,
657
],
"spans": [
{
"bbox": [
106,
644,
505,
657
],
"score": 1.0,
"content": "with targeted scenario, and a UAP designed specifically for EEG (Liu et al., 2021). For both FGSM",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
654,
506,
668
],
"spans": [
{
"bbox": [
105,
654,
390,
668
],
"score": 1.0,
"content": "and PGD, the ASR increases together with the maximum amplitude",
"type": "text"
},
{
"bbox": [
390,
657,
396,
665
],
"score": 0.65,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
397,
654,
506,
668
],
"score": 1.0,
"content": "of the perturbation. They",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
665,
505,
678
],
"spans": [
{
"bbox": [
106,
665,
505,
678
],
"score": 1.0,
"content": "always outperform the random noise, with PGD performing slightly better than FGSM. They reach",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
106,
677,
505,
689
],
"spans": [
{
"bbox": [
106,
677,
194,
689
],
"score": 1.0,
"content": "the maximum ASR of",
"type": "text"
},
{
"bbox": [
195,
677,
226,
687
],
"score": 0.88,
"content": "9 9 . 9 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
226,
677,
247,
689
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
247,
677,
273,
687
],
"score": 0.59,
"content": "1 0 \\mathrm { m V } .",
"type": "inline_equation"
},
{
"bbox": [
273,
677,
473,
689
],
"score": 1.0,
"content": "The post-attack classification accuracy drops from",
"type": "text"
},
{
"bbox": [
473,
677,
505,
687
],
"score": 0.86,
"content": "7 4 . 7 8 \\%",
"type": "inline_equation"
}
],
"index": 40
},
{
"bbox": [
105,
687,
505,
700
],
"spans": [
{
"bbox": [
105,
687,
116,
700
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
117,
688,
136,
698
],
"score": 0.86,
"content": "48 \\%",
"type": "inline_equation"
},
{
"bbox": [
136,
687,
259,
700
],
"score": 1.0,
"content": "for a perturbation amplitude of",
"type": "text"
},
{
"bbox": [
260,
688,
283,
698
],
"score": 0.64,
"content": "2 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
283,
687,
310,
700
],
"score": 1.0,
"content": "and to",
"type": "text"
},
{
"bbox": [
310,
688,
329,
698
],
"score": 0.87,
"content": "33 \\%",
"type": "inline_equation"
},
{
"bbox": [
330,
687,
344,
700
],
"score": 1.0,
"content": "for",
"type": "text"
},
{
"bbox": [
345,
688,
372,
698
],
"score": 0.75,
"content": "1 0 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
372,
687,
505,
700
],
"score": 1.0,
"content": "and higher amplitudes. Figure 2a",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
698,
507,
712
],
"spans": [
{
"bbox": [
105,
698,
507,
712
],
"score": 1.0,
"content": "shows the signals of a successful attack using PGD. The adversarial perturbation has a square-",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "wave form which negatively affects the natural shape of the EEG signal. By adding the proposed",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
106,
721,
505,
733
],
"spans": [
{
"bbox": [
106,
721,
505,
733
],
"score": 1.0,
"content": "derivative term, the square-wave artifacts are significantly reduced (2b), making the perturbation more",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
106,
83,
505,
95
],
"spans": [
{
"bbox": [
106,
83,
505,
95
],
"score": 1.0,
"content": "physiologically plausible. When comparing the power spectral density of the original and attacked",
"type": "text",
"cross_page": true
}
],
"index": 0
},
{
"bbox": [
105,
93,
506,
107
],
"spans": [
{
"bbox": [
105,
93,
506,
107
],
"score": 1.0,
"content": "signals, the attacked signal designed without derivative presents large components in low frequencies,",
"type": "text",
"cross_page": true
}
],
"index": 1
},
{
"bbox": [
105,
104,
506,
118
],
"spans": [
{
"bbox": [
105,
104,
506,
118
],
"score": 1.0,
"content": "making it more easily detectable. Whereas the attack with derivative loss better resembles the power",
"type": "text",
"cross_page": true
}
],
"index": 2
},
{
"bbox": [
105,
115,
506,
128
],
"spans": [
{
"bbox": [
105,
115,
506,
128
],
"score": 1.0,
"content": "spectral density of the original signal (see Appendix D). Moreover, the introduction of the derivative",
"type": "text",
"cross_page": true
}
],
"index": 3
},
{
"bbox": [
106,
126,
505,
139
],
"spans": [
{
"bbox": [
106,
126,
505,
139
],
"score": 1.0,
"content": "term does not degrade the ASR (Figure 1). The quantitative measures between the original and the",
"type": "text",
"cross_page": true
}
],
"index": 4
},
{
"bbox": [
105,
137,
506,
150
],
"spans": [
{
"bbox": [
105,
137,
506,
150
],
"score": 1.0,
"content": "adversarial samples in Table 1 demonstrate that our proposed method with derivative term generates",
"type": "text",
"cross_page": true
}
],
"index": 5
},
{
"bbox": [
106,
149,
505,
161
],
"spans": [
{
"bbox": [
106,
149,
505,
161
],
"score": 1.0,
"content": "adversarial samples that are more similar to the original EEG, allowing them to remain imperceptible",
"type": "text",
"cross_page": true
}
],
"index": 6
},
{
"bbox": [
105,
159,
506,
172
],
"spans": [
{
"bbox": [
105,
159,
168,
172
],
"score": 1.0,
"content": "even with high",
"type": "text",
"cross_page": true
},
{
"bbox": [
168,
161,
174,
169
],
"score": 0.56,
"content": "\\epsilon",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
174,
159,
506,
172
],
"score": 1.0,
"content": "(Appendix D). We reproduce the attacks using a Gaussian kernel as in (Han et al.,",
"type": "text",
"cross_page": true
}
],
"index": 7
},
{
"bbox": [
105,
169,
506,
184
],
"spans": [
{
"bbox": [
105,
169,
506,
184
],
"score": 1.0,
"content": "2020). After tuning the kernel size and variance of the Gaussian kernel, the method could not improve",
"type": "text",
"cross_page": true
}
],
"index": 8
},
{
"bbox": [
106,
181,
506,
194
],
"spans": [
{
"bbox": [
106,
181,
506,
194
],
"score": 1.0,
"content": "the plausibility metrics. The inferior performance of the Gaussian kernel could stem from the different",
"type": "text",
"cross_page": true
}
],
"index": 9
},
{
"bbox": [
105,
191,
496,
206
],
"spans": [
{
"bbox": [
105,
191,
496,
206
],
"score": 1.0,
"content": "nature of the signal: it was originally designed for ECGs which have a pseudo-periodic structure.",
"type": "text",
"cross_page": true
}
],
"index": 10
}
],
"index": 39,
"bbox_fs": [
105,
609,
507,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
203
],
"lines": [
{
"bbox": [
106,
83,
505,
95
],
"spans": [
{
"bbox": [
106,
83,
505,
95
],
"score": 1.0,
"content": "physiologically plausible. When comparing the power spectral density of the original and attacked",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
93,
506,
107
],
"spans": [
{
"bbox": [
105,
93,
506,
107
],
"score": 1.0,
"content": "signals, the attacked signal designed without derivative presents large components in low frequencies,",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
104,
506,
118
],
"spans": [
{
"bbox": [
105,
104,
506,
118
],
"score": 1.0,
"content": "making it more easily detectable. Whereas the attack with derivative loss better resembles the power",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
115,
506,
128
],
"spans": [
{
"bbox": [
105,
115,
506,
128
],
"score": 1.0,
"content": "spectral density of the original signal (see Appendix D). Moreover, the introduction of the derivative",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
126,
505,
139
],
"spans": [
{
"bbox": [
106,
126,
505,
139
],
"score": 1.0,
"content": "term does not degrade the ASR (Figure 1). The quantitative measures between the original and the",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
137,
506,
150
],
"spans": [
{
"bbox": [
105,
137,
506,
150
],
"score": 1.0,
"content": "adversarial samples in Table 1 demonstrate that our proposed method with derivative term generates",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
149,
505,
161
],
"spans": [
{
"bbox": [
106,
149,
505,
161
],
"score": 1.0,
"content": "adversarial samples that are more similar to the original EEG, allowing them to remain imperceptible",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
159,
506,
172
],
"spans": [
{
"bbox": [
105,
159,
168,
172
],
"score": 1.0,
"content": "even with high",
"type": "text"
},
{
"bbox": [
168,
161,
174,
169
],
"score": 0.56,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
174,
159,
506,
172
],
"score": 1.0,
"content": "(Appendix D). We reproduce the attacks using a Gaussian kernel as in (Han et al.,",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
169,
506,
184
],
"spans": [
{
"bbox": [
105,
169,
506,
184
],
"score": 1.0,
"content": "2020). After tuning the kernel size and variance of the Gaussian kernel, the method could not improve",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
181,
506,
194
],
"spans": [
{
"bbox": [
106,
181,
506,
194
],
"score": 1.0,
"content": "the plausibility metrics. The inferior performance of the Gaussian kernel could stem from the different",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
191,
496,
206
],
"spans": [
{
"bbox": [
105,
191,
496,
206
],
"score": 1.0,
"content": "nature of the signal: it was originally designed for ECGs which have a pseudo-periodic structure.",
"type": "text"
}
],
"index": 10
}
],
"index": 5
},
{
"type": "text",
"bbox": [
107,
209,
505,
275
],
"lines": [
{
"bbox": [
106,
209,
505,
221
],
"spans": [
{
"bbox": [
106,
209,
505,
221
],
"score": 1.0,
"content": "We extend the application of the derivative term to the UAP attack, while still not considering the",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
104,
219,
503,
233
],
"spans": [
{
"bbox": [
104,
219,
497,
233
],
"score": 1.0,
"content": "propagation model (Case 1). Figure 1 shows a comparison in performance for different values of",
"type": "text"
},
{
"bbox": [
498,
222,
503,
230
],
"score": 0.59,
"content": "\\epsilon",
"type": "inline_equation"
}
],
"index": 12
},
{
"bbox": [
105,
230,
505,
243
],
"spans": [
{
"bbox": [
105,
230,
290,
243
],
"score": 1.0,
"content": "The saturation in ASR is reached with higher",
"type": "text"
},
{
"bbox": [
290,
233,
296,
241
],
"score": 0.59,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
296,
230,
316,
243
],
"score": 1.0,
"content": ", i.e.,",
"type": "text"
},
{
"bbox": [
317,
231,
349,
241
],
"score": 0.88,
"content": "9 9 . 9 4 \\%",
"type": "inline_equation"
},
{
"bbox": [
349,
230,
370,
243
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
370,
231,
397,
241
],
"score": 0.64,
"content": "5 0 \\mathrm { m V } .",
"type": "inline_equation"
},
{
"bbox": [
397,
230,
505,
243
],
"score": 1.0,
"content": ". This is expected since the",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
241,
506,
254
],
"spans": [
{
"bbox": [
105,
241,
506,
254
],
"score": 1.0,
"content": "UAP is a more difficult attack where a single set of perturbations per EEG channel is generated for",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
252,
506,
265
],
"spans": [
{
"bbox": [
105,
252,
506,
265
],
"score": 1.0,
"content": "all the test samples. Likewise in PGD, the ASR does not drop with the addition of the derivative term.",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
263,
500,
276
],
"spans": [
{
"bbox": [
105,
263,
500,
276
],
"score": 1.0,
"content": "We reproduce the UAP proposed by (Liu et al., 2021). Our UAP consistently reaches higher ASR.",
"type": "text"
}
],
"index": 16
}
],
"index": 13.5
},
{
"type": "text",
"bbox": [
107,
290,
505,
444
],
"lines": [
{
"bbox": [
106,
290,
505,
303
],
"spans": [
{
"bbox": [
106,
290,
505,
303
],
"score": 1.0,
"content": "Spatial Propagation. Finally, we introduce the spatial constraints in the signal propagation over the",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
301,
505,
314
],
"spans": [
{
"bbox": [
106,
301,
505,
314
],
"score": 1.0,
"content": "scalp (Case 2). We consider 9 different scenarios by combining 3 realistic attenuation configurations",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
107,
309,
506,
327
],
"spans": [
{
"bbox": [
107,
312,
177,
324
],
"score": 0.91,
"content": "\\lambda _ { m } \\ \\in \\ \\{ 1 , 5 , 1 5 \\}",
"type": "inline_equation"
},
{
"bbox": [
178,
309,
299,
327
],
"score": 1.0,
"content": "with 3 delay configurations",
"type": "text"
},
{
"bbox": [
299,
312,
395,
324
],
"score": 0.93,
"content": "\\lambda _ { d } \\ \\in \\ \\{ 0 . 1 , 0 . 3 , 0 . 5 6 3 \\}",
"type": "inline_equation"
},
{
"bbox": [
396,
309,
506,
327
],
"score": 1.0,
"content": ", which capture the range",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
322,
505,
336
],
"spans": [
{
"bbox": [
105,
322,
505,
336
],
"score": 1.0,
"content": "described in Section 3.2. For evaluating the highest achievable attack efficiency, we test a scenario",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
334,
506,
347
],
"spans": [
{
"bbox": [
106,
334,
506,
347
],
"score": 1.0,
"content": "where the attacker is assumed to know the propagation model: the adversarial perturbation is generated",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
344,
505,
358
],
"spans": [
{
"bbox": [
105,
344,
272,
358
],
"score": 1.0,
"content": "and evaluated on fixed spatial parameters",
"type": "text"
},
{
"bbox": [
272,
345,
287,
356
],
"score": 0.9,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
287,
344,
304,
358
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
305,
345,
317,
356
],
"score": 0.88,
"content": "\\lambda _ { d }",
"type": "inline_equation"
},
{
"bbox": [
317,
344,
505,
358
],
"score": 1.0,
"content": ", shown in Figure 4, where the ASR reaches up",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
355,
506,
368
],
"spans": [
{
"bbox": [
106,
355,
116,
368
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
117,
356,
143,
366
],
"score": 0.87,
"content": "6 9 . 2 \\%",
"type": "inline_equation"
},
{
"bbox": [
144,
355,
202,
368
],
"score": 1.0,
"content": "with PGD and",
"type": "text"
},
{
"bbox": [
202,
356,
228,
366
],
"score": 0.86,
"content": "4 5 . 6 \\%",
"type": "inline_equation"
},
{
"bbox": [
229,
355,
280,
368
],
"score": 1.0,
"content": "with UAP at",
"type": "text"
},
{
"bbox": [
280,
356,
306,
366
],
"score": 0.61,
"content": "5 0 \\mathrm { m V } .",
"type": "inline_equation"
},
{
"bbox": [
306,
355,
506,
368
],
"score": 1.0,
"content": "Figure 3 depicts an example of a successful attack",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
367,
506,
380
],
"spans": [
{
"bbox": [
106,
367,
506,
380
],
"score": 1.0,
"content": "with the highest perturbation amplitude. The introduction of the spatial constraints makes the attack",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
377,
506,
390
],
"spans": [
{
"bbox": [
105,
377,
506,
390
],
"score": 1.0,
"content": "problem harder yielding seldom square distortions. However, the resulting EEG signals still resemble",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
388,
506,
403
],
"spans": [
{
"bbox": [
105,
388,
506,
403
],
"score": 1.0,
"content": "physiological random processes typical of EEGs. Next, we ablate the spatial constraints during",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
399,
506,
413
],
"spans": [
{
"bbox": [
105,
399,
506,
413
],
"score": 1.0,
"content": "generation and test the resulting perturbations on the 9 above-mentioned scenarios (Case 3). The",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
410,
505,
423
],
"spans": [
{
"bbox": [
105,
410,
263,
423
],
"score": 1.0,
"content": "ASR drops significantly, especially for",
"type": "text"
},
{
"bbox": [
264,
411,
289,
422
],
"score": 0.91,
"content": "\\lambda _ { m } { = } 5",
"type": "inline_equation"
},
{
"bbox": [
290,
410,
307,
423
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
307,
411,
338,
422
],
"score": 0.91,
"content": "\\lambda _ { m } { = } 1 5",
"type": "inline_equation"
},
{
"bbox": [
338,
410,
505,
423
],
"score": 1.0,
"content": "where the attenuation of the perturbation",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
421,
506,
434
],
"spans": [
{
"bbox": [
105,
421,
506,
434
],
"score": 1.0,
"content": "over the scalp is greater (see Figure 7), and with the global UAP attack, where the attacker does not",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
433,
270,
445
],
"spans": [
{
"bbox": [
106,
433,
270,
445
],
"score": 1.0,
"content": "have access to the attacked EEG signals.",
"type": "text"
}
],
"index": 30
}
],
"index": 23.5
},
{
"type": "text",
"bbox": [
107,
450,
505,
538
],
"lines": [
{
"bbox": [
106,
450,
506,
462
],
"spans": [
{
"bbox": [
106,
450,
506,
462
],
"score": 1.0,
"content": "Our spatial propagation models allows us to identify the vulnerability of the individual EEG channels.",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
460,
505,
473
],
"spans": [
{
"bbox": [
105,
460,
505,
473
],
"score": 1.0,
"content": "Figure 5 shows the ASR when initiating an attack from a specific channel (T9, T10, etc.) and",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
471,
506,
484
],
"spans": [
{
"bbox": [
105,
471,
421,
484
],
"score": 1.0,
"content": "propagating it to the rest of the head. In the case with the greatest attenuation",
"type": "text"
},
{
"bbox": [
421,
472,
455,
483
],
"score": 0.86,
"content": "\\left( \\lambda _ { m } = 1 5 \\right)",
"type": "inline_equation"
},
{
"bbox": [
456,
471,
506,
484
],
"score": 1.0,
"content": "we find the",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
482,
505,
495
],
"spans": [
{
"bbox": [
105,
482,
235,
495
],
"score": 1.0,
"content": "maximum ASR at the electrode",
"type": "text"
},
{
"bbox": [
235,
483,
248,
493
],
"score": 0.6,
"content": "\\mathbf { C } \\mathbf { z }",
"type": "inline_equation"
},
{
"bbox": [
249,
482,
505,
495
],
"score": 1.0,
"content": "between the regions of the electrodes C3 and C4, which are the",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
493,
505,
505
],
"spans": [
{
"bbox": [
106,
493,
505,
505
],
"score": 1.0,
"content": "most relevant ones for MI of the left and right hand tasks (Pfurtscheller & Lopes da Silva, 1999). We",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
504,
506,
516
],
"spans": [
{
"bbox": [
105,
504,
506,
516
],
"score": 1.0,
"content": "compute the pre- and post-attack confusion matrices for attacks from T9 and T10 (see Appendix E).",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
515,
505,
528
],
"spans": [
{
"bbox": [
105,
515,
505,
528
],
"score": 1.0,
"content": "When the attack propagates from the left side (T9), more samples with ground-truth label “right” can",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
527,
483,
538
],
"spans": [
{
"bbox": [
106,
527,
483,
538
],
"score": 1.0,
"content": "be fooled to “rest”, while the attacks from the right side (T10) are more effective “left” labels.",
"type": "text"
}
],
"index": 38
}
],
"index": 34.5
},
{
"type": "text",
"bbox": [
108,
543,
505,
587
],
"lines": [
{
"bbox": [
105,
542,
505,
556
],
"spans": [
{
"bbox": [
105,
542,
505,
556
],
"score": 1.0,
"content": "Overall, our methods successfully generates perturbations resembling natural noise in EEGs, that can",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
554,
506,
567
],
"spans": [
{
"bbox": [
105,
554,
506,
567
],
"score": 1.0,
"content": "be added at the source of the signal acquisition and are propagated over the scalp, creating attacked",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
566,
505,
578
],
"spans": [
{
"bbox": [
106,
566,
505,
578
],
"score": 1.0,
"content": "signals that are physiologically plausible. Similar results have been observed on the BCI Competition",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
576,
255,
587
],
"spans": [
{
"bbox": [
105,
576,
255,
587
],
"score": 1.0,
"content": "IV-2a dataset, shown in Appendix C.",
"type": "text"
}
],
"index": 42
}
],
"index": 40.5
},
{
"type": "title",
"bbox": [
108,
606,
195,
619
],
"lines": [
{
"bbox": [
104,
605,
197,
622
],
"spans": [
{
"bbox": [
104,
605,
197,
622
],
"score": 1.0,
"content": "5 CONCLUSION",
"type": "text"
}
],
"index": 43
}
],
"index": 43
},
{
"type": "text",
"bbox": [
107,
632,
506,
732
],
"lines": [
{
"bbox": [
106,
633,
505,
645
],
"spans": [
{
"bbox": [
106,
633,
505,
645
],
"score": 1.0,
"content": "With the incentive of improving security in BCIs, in this work, we demonstrated that DoS attacks",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
106,
644,
505,
656
],
"spans": [
{
"bbox": [
106,
644,
505,
656
],
"score": 1.0,
"content": "are feasible and effective despite physical domain constraints. Experimental results reveal potential",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
106,
655,
506,
667
],
"spans": [
{
"bbox": [
106,
655,
506,
667
],
"score": 1.0,
"content": "risks of realistic attacks on smart wearable BCIs and incentivize the need for future development of",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
665,
506,
678
],
"spans": [
{
"bbox": [
105,
665,
506,
678
],
"score": 1.0,
"content": "defense mechanisms while designing deep learning models to be embedded in smart wearable BCIs.",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
676,
506,
690
],
"spans": [
{
"bbox": [
105,
676,
506,
690
],
"score": 1.0,
"content": "Our detailed analysis on each EEG channel shows that special attention has to be paid, combined",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
106,
688,
506,
700
],
"spans": [
{
"bbox": [
106,
688,
506,
700
],
"score": 1.0,
"content": "with the findings in neuroscience, to the brain regions that are found responsible for a specific task.",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
699,
506,
712
],
"spans": [
{
"bbox": [
105,
699,
506,
712
],
"score": 1.0,
"content": "In future work, the proposed attacks can cover uncertainty in the propagation model and the timing of",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
106,
710,
505,
722
],
"spans": [
{
"bbox": [
106,
710,
505,
722
],
"score": 1.0,
"content": "the MI activity. Moreover, hardware implementations of such attacks can be created to evaluate the",
"type": "text"
}
],
"index": 51
},
{
"bbox": [
105,
721,
494,
733
],
"spans": [
{
"bbox": [
105,
721,
494,
733
],
"score": 1.0,
"content": "proposed methods in real-world, with the ultimate goal of developing effective countermeasures.",
"type": "text"
}
],
"index": 52
}
],
"index": 48
}
],
"page_idx": 8,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
309,
39
],
"spans": [
{
"bbox": [
106,
25,
309,
39
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
308,
759
],
"lines": [
{
"bbox": [
302,
751,
309,
762
],
"spans": [
{
"bbox": [
302,
751,
309,
762
],
"score": 1.0,
"content": "9",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
203
],
"lines": [],
"index": 5,
"bbox_fs": [
105,
83,
506,
206
],
"lines_deleted": true
},
{
"type": "text",
"bbox": [
107,
209,
505,
275
],
"lines": [
{
"bbox": [
106,
209,
505,
221
],
"spans": [
{
"bbox": [
106,
209,
505,
221
],
"score": 1.0,
"content": "We extend the application of the derivative term to the UAP attack, while still not considering the",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
104,
219,
503,
233
],
"spans": [
{
"bbox": [
104,
219,
497,
233
],
"score": 1.0,
"content": "propagation model (Case 1). Figure 1 shows a comparison in performance for different values of",
"type": "text"
},
{
"bbox": [
498,
222,
503,
230
],
"score": 0.59,
"content": "\\epsilon",
"type": "inline_equation"
}
],
"index": 12
},
{
"bbox": [
105,
230,
505,
243
],
"spans": [
{
"bbox": [
105,
230,
290,
243
],
"score": 1.0,
"content": "The saturation in ASR is reached with higher",
"type": "text"
},
{
"bbox": [
290,
233,
296,
241
],
"score": 0.59,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
296,
230,
316,
243
],
"score": 1.0,
"content": ", i.e.,",
"type": "text"
},
{
"bbox": [
317,
231,
349,
241
],
"score": 0.88,
"content": "9 9 . 9 4 \\%",
"type": "inline_equation"
},
{
"bbox": [
349,
230,
370,
243
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
370,
231,
397,
241
],
"score": 0.64,
"content": "5 0 \\mathrm { m V } .",
"type": "inline_equation"
},
{
"bbox": [
397,
230,
505,
243
],
"score": 1.0,
"content": ". This is expected since the",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
241,
506,
254
],
"spans": [
{
"bbox": [
105,
241,
506,
254
],
"score": 1.0,
"content": "UAP is a more difficult attack where a single set of perturbations per EEG channel is generated for",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
252,
506,
265
],
"spans": [
{
"bbox": [
105,
252,
506,
265
],
"score": 1.0,
"content": "all the test samples. Likewise in PGD, the ASR does not drop with the addition of the derivative term.",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
263,
500,
276
],
"spans": [
{
"bbox": [
105,
263,
500,
276
],
"score": 1.0,
"content": "We reproduce the UAP proposed by (Liu et al., 2021). Our UAP consistently reaches higher ASR.",
"type": "text"
}
],
"index": 16
}
],
"index": 13.5,
"bbox_fs": [
104,
209,
506,
276
]
},
{
"type": "text",
"bbox": [
107,
290,
505,
444
],
"lines": [
{
"bbox": [
106,
290,
505,
303
],
"spans": [
{
"bbox": [
106,
290,
505,
303
],
"score": 1.0,
"content": "Spatial Propagation. Finally, we introduce the spatial constraints in the signal propagation over the",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
301,
505,
314
],
"spans": [
{
"bbox": [
106,
301,
505,
314
],
"score": 1.0,
"content": "scalp (Case 2). We consider 9 different scenarios by combining 3 realistic attenuation configurations",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
107,
309,
506,
327
],
"spans": [
{
"bbox": [
107,
312,
177,
324
],
"score": 0.91,
"content": "\\lambda _ { m } \\ \\in \\ \\{ 1 , 5 , 1 5 \\}",
"type": "inline_equation"
},
{
"bbox": [
178,
309,
299,
327
],
"score": 1.0,
"content": "with 3 delay configurations",
"type": "text"
},
{
"bbox": [
299,
312,
395,
324
],
"score": 0.93,
"content": "\\lambda _ { d } \\ \\in \\ \\{ 0 . 1 , 0 . 3 , 0 . 5 6 3 \\}",
"type": "inline_equation"
},
{
"bbox": [
396,
309,
506,
327
],
"score": 1.0,
"content": ", which capture the range",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
322,
505,
336
],
"spans": [
{
"bbox": [
105,
322,
505,
336
],
"score": 1.0,
"content": "described in Section 3.2. For evaluating the highest achievable attack efficiency, we test a scenario",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
334,
506,
347
],
"spans": [
{
"bbox": [
106,
334,
506,
347
],
"score": 1.0,
"content": "where the attacker is assumed to know the propagation model: the adversarial perturbation is generated",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
344,
505,
358
],
"spans": [
{
"bbox": [
105,
344,
272,
358
],
"score": 1.0,
"content": "and evaluated on fixed spatial parameters",
"type": "text"
},
{
"bbox": [
272,
345,
287,
356
],
"score": 0.9,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
287,
344,
304,
358
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
305,
345,
317,
356
],
"score": 0.88,
"content": "\\lambda _ { d }",
"type": "inline_equation"
},
{
"bbox": [
317,
344,
505,
358
],
"score": 1.0,
"content": ", shown in Figure 4, where the ASR reaches up",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
355,
506,
368
],
"spans": [
{
"bbox": [
106,
355,
116,
368
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
117,
356,
143,
366
],
"score": 0.87,
"content": "6 9 . 2 \\%",
"type": "inline_equation"
},
{
"bbox": [
144,
355,
202,
368
],
"score": 1.0,
"content": "with PGD and",
"type": "text"
},
{
"bbox": [
202,
356,
228,
366
],
"score": 0.86,
"content": "4 5 . 6 \\%",
"type": "inline_equation"
},
{
"bbox": [
229,
355,
280,
368
],
"score": 1.0,
"content": "with UAP at",
"type": "text"
},
{
"bbox": [
280,
356,
306,
366
],
"score": 0.61,
"content": "5 0 \\mathrm { m V } .",
"type": "inline_equation"
},
{
"bbox": [
306,
355,
506,
368
],
"score": 1.0,
"content": "Figure 3 depicts an example of a successful attack",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
367,
506,
380
],
"spans": [
{
"bbox": [
106,
367,
506,
380
],
"score": 1.0,
"content": "with the highest perturbation amplitude. The introduction of the spatial constraints makes the attack",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
377,
506,
390
],
"spans": [
{
"bbox": [
105,
377,
506,
390
],
"score": 1.0,
"content": "problem harder yielding seldom square distortions. However, the resulting EEG signals still resemble",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
388,
506,
403
],
"spans": [
{
"bbox": [
105,
388,
506,
403
],
"score": 1.0,
"content": "physiological random processes typical of EEGs. Next, we ablate the spatial constraints during",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
399,
506,
413
],
"spans": [
{
"bbox": [
105,
399,
506,
413
],
"score": 1.0,
"content": "generation and test the resulting perturbations on the 9 above-mentioned scenarios (Case 3). The",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
410,
505,
423
],
"spans": [
{
"bbox": [
105,
410,
263,
423
],
"score": 1.0,
"content": "ASR drops significantly, especially for",
"type": "text"
},
{
"bbox": [
264,
411,
289,
422
],
"score": 0.91,
"content": "\\lambda _ { m } { = } 5",
"type": "inline_equation"
},
{
"bbox": [
290,
410,
307,
423
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
307,
411,
338,
422
],
"score": 0.91,
"content": "\\lambda _ { m } { = } 1 5",
"type": "inline_equation"
},
{
"bbox": [
338,
410,
505,
423
],
"score": 1.0,
"content": "where the attenuation of the perturbation",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
421,
506,
434
],
"spans": [
{
"bbox": [
105,
421,
506,
434
],
"score": 1.0,
"content": "over the scalp is greater (see Figure 7), and with the global UAP attack, where the attacker does not",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
433,
270,
445
],
"spans": [
{
"bbox": [
106,
433,
270,
445
],
"score": 1.0,
"content": "have access to the attacked EEG signals.",
"type": "text"
}
],
"index": 30
}
],
"index": 23.5,
"bbox_fs": [
105,
290,
506,
445
]
},
{
"type": "text",
"bbox": [
107,
450,
505,
538
],
"lines": [
{
"bbox": [
106,
450,
506,
462
],
"spans": [
{
"bbox": [
106,
450,
506,
462
],
"score": 1.0,
"content": "Our spatial propagation models allows us to identify the vulnerability of the individual EEG channels.",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
460,
505,
473
],
"spans": [
{
"bbox": [
105,
460,
505,
473
],
"score": 1.0,
"content": "Figure 5 shows the ASR when initiating an attack from a specific channel (T9, T10, etc.) and",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
471,
506,
484
],
"spans": [
{
"bbox": [
105,
471,
421,
484
],
"score": 1.0,
"content": "propagating it to the rest of the head. In the case with the greatest attenuation",
"type": "text"
},
{
"bbox": [
421,
472,
455,
483
],
"score": 0.86,
"content": "\\left( \\lambda _ { m } = 1 5 \\right)",
"type": "inline_equation"
},
{
"bbox": [
456,
471,
506,
484
],
"score": 1.0,
"content": "we find the",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
482,
505,
495
],
"spans": [
{
"bbox": [
105,
482,
235,
495
],
"score": 1.0,
"content": "maximum ASR at the electrode",
"type": "text"
},
{
"bbox": [
235,
483,
248,
493
],
"score": 0.6,
"content": "\\mathbf { C } \\mathbf { z }",
"type": "inline_equation"
},
{
"bbox": [
249,
482,
505,
495
],
"score": 1.0,
"content": "between the regions of the electrodes C3 and C4, which are the",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
493,
505,
505
],
"spans": [
{
"bbox": [
106,
493,
505,
505
],
"score": 1.0,
"content": "most relevant ones for MI of the left and right hand tasks (Pfurtscheller & Lopes da Silva, 1999). We",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
504,
506,
516
],
"spans": [
{
"bbox": [
105,
504,
506,
516
],
"score": 1.0,
"content": "compute the pre- and post-attack confusion matrices for attacks from T9 and T10 (see Appendix E).",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
515,
505,
528
],
"spans": [
{
"bbox": [
105,
515,
505,
528
],
"score": 1.0,
"content": "When the attack propagates from the left side (T9), more samples with ground-truth label “right” can",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
527,
483,
538
],
"spans": [
{
"bbox": [
106,
527,
483,
538
],
"score": 1.0,
"content": "be fooled to “rest”, while the attacks from the right side (T10) are more effective “left” labels.",
"type": "text"
}
],
"index": 38
}
],
"index": 34.5,
"bbox_fs": [
105,
450,
506,
538
]
},
{
"type": "text",
"bbox": [
108,
543,
505,
587
],
"lines": [
{
"bbox": [
105,
542,
505,
556
],
"spans": [
{
"bbox": [
105,
542,
505,
556
],
"score": 1.0,
"content": "Overall, our methods successfully generates perturbations resembling natural noise in EEGs, that can",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
554,
506,
567
],
"spans": [
{
"bbox": [
105,
554,
506,
567
],
"score": 1.0,
"content": "be added at the source of the signal acquisition and are propagated over the scalp, creating attacked",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
566,
505,
578
],
"spans": [
{
"bbox": [
106,
566,
505,
578
],
"score": 1.0,
"content": "signals that are physiologically plausible. Similar results have been observed on the BCI Competition",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
576,
255,
587
],
"spans": [
{
"bbox": [
105,
576,
255,
587
],
"score": 1.0,
"content": "IV-2a dataset, shown in Appendix C.",
"type": "text"
}
],
"index": 42
}
],
"index": 40.5,
"bbox_fs": [
105,
542,
506,
587
]
},
{
"type": "title",
"bbox": [
108,
606,
195,
619
],
"lines": [
{
"bbox": [
104,
605,
197,
622
],
"spans": [
{
"bbox": [
104,
605,
197,
622
],
"score": 1.0,
"content": "5 CONCLUSION",
"type": "text"
}
],
"index": 43
}
],
"index": 43
},
{
"type": "text",
"bbox": [
107,
632,
506,
732
],
"lines": [
{
"bbox": [
106,
633,
505,
645
],
"spans": [
{
"bbox": [
106,
633,
505,
645
],
"score": 1.0,
"content": "With the incentive of improving security in BCIs, in this work, we demonstrated that DoS attacks",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
106,
644,
505,
656
],
"spans": [
{
"bbox": [
106,
644,
505,
656
],
"score": 1.0,
"content": "are feasible and effective despite physical domain constraints. Experimental results reveal potential",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
106,
655,
506,
667
],
"spans": [
{
"bbox": [
106,
655,
506,
667
],
"score": 1.0,
"content": "risks of realistic attacks on smart wearable BCIs and incentivize the need for future development of",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
665,
506,
678
],
"spans": [
{
"bbox": [
105,
665,
506,
678
],
"score": 1.0,
"content": "defense mechanisms while designing deep learning models to be embedded in smart wearable BCIs.",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
105,
676,
506,
690
],
"spans": [
{
"bbox": [
105,
676,
506,
690
],
"score": 1.0,
"content": "Our detailed analysis on each EEG channel shows that special attention has to be paid, combined",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
106,
688,
506,
700
],
"spans": [
{
"bbox": [
106,
688,
506,
700
],
"score": 1.0,
"content": "with the findings in neuroscience, to the brain regions that are found responsible for a specific task.",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
699,
506,
712
],
"spans": [
{
"bbox": [
105,
699,
506,
712
],
"score": 1.0,
"content": "In future work, the proposed attacks can cover uncertainty in the propagation model and the timing of",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
106,
710,
505,
722
],
"spans": [
{
"bbox": [
106,
710,
505,
722
],
"score": 1.0,
"content": "the MI activity. Moreover, hardware implementations of such attacks can be created to evaluate the",
"type": "text"
}
],
"index": 51
},
{
"bbox": [
105,
721,
494,
733
],
"spans": [
{
"bbox": [
105,
721,
494,
733
],
"score": 1.0,
"content": "proposed methods in real-world, with the ultimate goal of developing effective countermeasures.",
"type": "text"
}
],
"index": 52
}
],
"index": 48,
"bbox_fs": [
105,
633,
506,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
107,
82,
194,
93
],
"lines": [
{
"bbox": [
106,
82,
195,
94
],
"spans": [
{
"bbox": [
106,
82,
195,
94
],
"score": 1.0,
"content": "ETHICS STATEMENT",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
107,
102,
505,
201
],
"lines": [
{
"bbox": [
105,
101,
506,
115
],
"spans": [
{
"bbox": [
105,
101,
506,
115
],
"score": 1.0,
"content": "The sore goal of this work is to raise awareness of potential adversarial attacks in BCIs and incentivize",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
113,
505,
125
],
"spans": [
{
"bbox": [
106,
113,
505,
125
],
"score": 1.0,
"content": "the development of coutermeasures, especially in the current moment when the BCIs are facing an",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
124,
505,
136
],
"spans": [
{
"bbox": [
105,
124,
505,
136
],
"score": 1.0,
"content": "increasing growth in applications of everyday life. The active development of smart wearable BCIs is",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
135,
507,
147
],
"spans": [
{
"bbox": [
105,
135,
507,
147
],
"score": 1.0,
"content": "introducing a paradigm shift where the processing algorithms are embedded near the data acquisition.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
146,
506,
159
],
"spans": [
{
"bbox": [
105,
146,
506,
159
],
"score": 1.0,
"content": "While this improves the system security to a certain extend, with this work we have shown that it is",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
156,
506,
169
],
"spans": [
{
"bbox": [
105,
156,
506,
169
],
"score": 1.0,
"content": "not the only and ultimate way to a safe and reliable BCI, since we have shown that BCI systems are",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
106,
168,
505,
180
],
"spans": [
{
"bbox": [
106,
168,
505,
180
],
"score": 1.0,
"content": "vulnerable also to attacks at the signals’ source. We hope that our work sheds light on the fact that",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
178,
506,
191
],
"spans": [
{
"bbox": [
105,
178,
506,
191
],
"score": 1.0,
"content": "practical BCI systems are vulnerable, despite the physical constraints, and motivates the design and",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
190,
325,
203
],
"spans": [
{
"bbox": [
105,
190,
325,
203
],
"score": 1.0,
"content": "development of more reliable and robust BCI systems.",
"type": "text"
}
],
"index": 9
}
],
"index": 5
},
{
"type": "title",
"bbox": [
108,
214,
184,
225
],
"lines": [
{
"bbox": [
106,
214,
185,
226
],
"spans": [
{
"bbox": [
106,
214,
185,
226
],
"score": 1.0,
"content": "REPRODUCIBLITY",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
108,
233,
504,
255
],
"lines": [
{
"bbox": [
104,
231,
506,
248
],
"spans": [
{
"bbox": [
104,
231,
506,
248
],
"score": 1.0,
"content": "A link to a anonymous downloadable source code of this work is submitted as supplementary",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
244,
149,
256
],
"spans": [
{
"bbox": [
106,
244,
149,
256
],
"score": 1.0,
"content": "materials.",
"type": "text"
}
],
"index": 12
}
],
"index": 11.5
},
{
"type": "title",
"bbox": [
108,
272,
175,
285
],
"lines": [
{
"bbox": [
106,
272,
176,
286
],
"spans": [
{
"bbox": [
106,
272,
176,
286
],
"score": 1.0,
"content": "REFERENCES",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
107,
292,
384,
303
],
"lines": [
{
"bbox": [
105,
290,
385,
306
],
"spans": [
{
"bbox": [
105,
290,
385,
306
],
"score": 1.0,
"content": "BCI2000: A general-purpose brain-computer interface (BCI) system.",
"type": "text"
}
],
"index": 14
}
],
"index": 14
},
{
"type": "text",
"bbox": [
104,
312,
505,
334
],
"lines": [
{
"bbox": [
105,
311,
506,
325
],
"spans": [
{
"bbox": [
105,
311,
506,
325
],
"score": 1.0,
"content": "Ayten Ozge Akmandor and Niraj K. Jha. Smart health care: An edge-side computing perspective.",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
115,
322,
349,
335
],
"spans": [
{
"bbox": [
115,
322,
349,
335
],
"score": 1.0,
"content": "IEEE Consumer Electronics Magazine, 7(1):29–37, 2018.",
"type": "text"
}
],
"index": 16
}
],
"index": 15.5
},
{
"type": "text",
"bbox": [
107,
342,
504,
366
],
"lines": [
{
"bbox": [
106,
342,
504,
355
],
"spans": [
{
"bbox": [
106,
342,
504,
355
],
"score": 1.0,
"content": "Ralph Algazi, Carlos Avendano, and Richard O Duda. Estimation of a spherical-head model from",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
116,
353,
293,
366
],
"spans": [
{
"bbox": [
116,
353,
293,
366
],
"score": 1.0,
"content": "anthropometry. J. Aud. Eng. Soc., 49, 2001.",
"type": "text"
}
],
"index": 18
}
],
"index": 17.5
},
{
"type": "text",
"bbox": [
106,
373,
503,
396
],
"lines": [
{
"bbox": [
105,
373,
505,
386
],
"spans": [
{
"bbox": [
105,
373,
505,
386
],
"score": 1.0,
"content": "Pietro Aricò, Nicolina Sciaraffa, and Fabio Babiloni. Brain–computer interfaces: Toward a daily life",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
115,
385,
288,
397
],
"spans": [
{
"bbox": [
115,
385,
288,
397
],
"score": 1.0,
"content": "employment. Brain Sciences, 10(3), 2020.",
"type": "text"
}
],
"index": 20
}
],
"index": 19.5
},
{
"type": "text",
"bbox": [
106,
404,
506,
439
],
"lines": [
{
"bbox": [
106,
405,
505,
417
],
"spans": [
{
"bbox": [
106,
405,
505,
417
],
"score": 1.0,
"content": "Aamir Arsalan, Muhammad Majid, Amna Rauf Butt, and Syed Muhammad Anwar. Classification",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
116,
415,
506,
429
],
"spans": [
{
"bbox": [
116,
415,
506,
429
],
"score": 1.0,
"content": "of perceived mental stress using a commercially available EEG headband. IEEE Journal of",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
116,
427,
361,
439
],
"spans": [
{
"bbox": [
116,
427,
361,
439
],
"score": 1.0,
"content": "Biomedical and Health Informatics, 23(6):2257–2264, 2019.",
"type": "text"
}
],
"index": 23
}
],
"index": 22
},
{
"type": "text",
"bbox": [
108,
446,
505,
481
],
"lines": [
{
"bbox": [
106,
446,
505,
459
],
"spans": [
{
"bbox": [
106,
446,
505,
459
],
"score": 1.0,
"content": "Christopher Beach, Ertan Balaban, and Alexander J. Casson. Chapter 14 - edge algorithms for",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
115,
458,
505,
470
],
"spans": [
{
"bbox": [
115,
458,
505,
470
],
"score": 1.0,
"content": "wearables: an overview of a truly multi-disciplinary problem. In Edward Sazonov (ed.), Wearable",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
115,
469,
497,
481
],
"spans": [
{
"bbox": [
115,
469,
497,
481
],
"score": 1.0,
"content": "Sensors (Second Edition), pp. 379–414. Academic Press, Oxford, second edition edition, 2021.",
"type": "text"
}
],
"index": 26
}
],
"index": 25
},
{
"type": "text",
"bbox": [
108,
488,
504,
523
],
"lines": [
{
"bbox": [
106,
488,
505,
501
],
"spans": [
{
"bbox": [
106,
488,
505,
501
],
"score": 1.0,
"content": "Kais Belwafi, Olivier Romain, Sofien Gannouni, Fakhreddine Ghaffari, Ridha Djemal, and Bouraoui",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
115,
500,
506,
513
],
"spans": [
{
"bbox": [
115,
500,
506,
513
],
"score": 1.0,
"content": "Ouni. An embedded implementation based on adaptive filter bank for brain–computer interface",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
115,
511,
318,
523
],
"spans": [
{
"bbox": [
115,
511,
318,
523
],
"score": 1.0,
"content": "systems. Journal of Neuroscience Methods, 2018.",
"type": "text"
}
],
"index": 29
}
],
"index": 28
},
{
"type": "text",
"bbox": [
107,
530,
506,
564
],
"lines": [
{
"bbox": [
105,
529,
506,
544
],
"spans": [
{
"bbox": [
105,
529,
506,
544
],
"score": 1.0,
"content": "Sergio López Bernal, Alberto Huertas Celdrán, Gregorio Martínez Pérez, Michael Taynnan Barros,",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
116,
542,
506,
554
],
"spans": [
{
"bbox": [
116,
542,
506,
554
],
"score": 1.0,
"content": "and Sasitharan Balasubramaniam. Security in brain-computer interfaces: State-of-the-art,",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
115,
552,
436,
565
],
"spans": [
{
"bbox": [
115,
552,
436,
565
],
"score": 1.0,
"content": "opportunities, and future challenges. ACM Comput. Surv., 54(1), January 2021.",
"type": "text"
}
],
"index": 32
}
],
"index": 31
},
{
"type": "text",
"bbox": [
106,
572,
506,
617
],
"lines": [
{
"bbox": [
105,
572,
506,
585
],
"spans": [
{
"bbox": [
105,
572,
506,
585
],
"score": 1.0,
"content": "Florian C. A. A. Bodranghien, Margot Langlois Mahe, Serge Clément, and Mario U. Manto. A pilot",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
115,
582,
506,
597
],
"spans": [
{
"bbox": [
115,
582,
506,
597
],
"score": 1.0,
"content": "study on the effects of transcranial direct current stimulation on brain rhythms and entropy during",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
115,
595,
507,
608
],
"spans": [
{
"bbox": [
115,
595,
507,
608
],
"score": 1.0,
"content": "self-paced finger movement using the Epoc helmet. Frontiers in Human Neuroscience, 11:201,",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
115,
604,
143,
617
],
"spans": [
{
"bbox": [
115,
604,
143,
617
],
"score": 1.0,
"content": "2017.",
"type": "text"
}
],
"index": 36
}
],
"index": 34.5
},
{
"type": "text",
"bbox": [
107,
625,
506,
670
],
"lines": [
{
"bbox": [
105,
626,
506,
639
],
"spans": [
{
"bbox": [
105,
626,
506,
639
],
"score": 1.0,
"content": "Jonathan S. Brumberg, Jeremy D. Burnison, and Kevin M. Pitt. Using motor imagery to control",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
115,
636,
506,
649
],
"spans": [
{
"bbox": [
115,
636,
506,
649
],
"score": 1.0,
"content": "brain-computer interfaces for communication. In Dylan D. Schmorrow and Cali M. Fidopiastis",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
115,
647,
507,
662
],
"spans": [
{
"bbox": [
115,
647,
507,
662
],
"score": 1.0,
"content": "(eds.), Foundations of Augmented Cognition: Neuroergonomics and Operational Neuroscience, pp.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
115,
658,
337,
672
],
"spans": [
{
"bbox": [
115,
658,
337,
672
],
"score": 1.0,
"content": "14–25, Cham, 2016. Springer International Publishing.",
"type": "text"
}
],
"index": 40
}
],
"index": 38.5
},
{
"type": "text",
"bbox": [
106,
678,
505,
701
],
"lines": [
{
"bbox": [
105,
678,
505,
691
],
"spans": [
{
"bbox": [
105,
678,
505,
691
],
"score": 1.0,
"content": "C Brunner, R Leeb, G R Müller-Putz, A Schlögl, and G Pfurtscheller. BCI competition 2008 - Graz",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
115,
688,
492,
702
],
"spans": [
{
"bbox": [
115,
688,
492,
702
],
"score": 1.0,
"content": "data set A, 2008. URL http://bnci-horizon-2020.eu/database/data-sets.",
"type": "text"
}
],
"index": 42
}
],
"index": 41.5
},
{
"type": "text",
"bbox": [
108,
709,
504,
732
],
"lines": [
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "Ujwal Chaudhary, Niels Birbaumer, and Ander Ramos-Murguialday. Brain–computer interfaces for",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
116,
720,
420,
733
],
"spans": [
{
"bbox": [
116,
720,
420,
733
],
"score": 1.0,
"content": "communication and rehabilitation. Nature Reviews Neurology, 12, 08 2016.",
"type": "text"
}
],
"index": 44
}
],
"index": 43.5
}
],
"page_idx": 9,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
308,
38
],
"spans": [
{
"bbox": [
106,
25,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
311,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
764
],
"spans": [
{
"bbox": [
299,
750,
312,
764
],
"score": 1.0,
"content": "10",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
107,
82,
194,
93
],
"lines": [
{
"bbox": [
106,
82,
195,
94
],
"spans": [
{
"bbox": [
106,
82,
195,
94
],
"score": 1.0,
"content": "ETHICS STATEMENT",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
107,
102,
505,
201
],
"lines": [
{
"bbox": [
105,
101,
506,
115
],
"spans": [
{
"bbox": [
105,
101,
506,
115
],
"score": 1.0,
"content": "The sore goal of this work is to raise awareness of potential adversarial attacks in BCIs and incentivize",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
113,
505,
125
],
"spans": [
{
"bbox": [
106,
113,
505,
125
],
"score": 1.0,
"content": "the development of coutermeasures, especially in the current moment when the BCIs are facing an",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
124,
505,
136
],
"spans": [
{
"bbox": [
105,
124,
505,
136
],
"score": 1.0,
"content": "increasing growth in applications of everyday life. The active development of smart wearable BCIs is",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
135,
507,
147
],
"spans": [
{
"bbox": [
105,
135,
507,
147
],
"score": 1.0,
"content": "introducing a paradigm shift where the processing algorithms are embedded near the data acquisition.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
146,
506,
159
],
"spans": [
{
"bbox": [
105,
146,
506,
159
],
"score": 1.0,
"content": "While this improves the system security to a certain extend, with this work we have shown that it is",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
156,
506,
169
],
"spans": [
{
"bbox": [
105,
156,
506,
169
],
"score": 1.0,
"content": "not the only and ultimate way to a safe and reliable BCI, since we have shown that BCI systems are",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
106,
168,
505,
180
],
"spans": [
{
"bbox": [
106,
168,
505,
180
],
"score": 1.0,
"content": "vulnerable also to attacks at the signals’ source. We hope that our work sheds light on the fact that",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
178,
506,
191
],
"spans": [
{
"bbox": [
105,
178,
506,
191
],
"score": 1.0,
"content": "practical BCI systems are vulnerable, despite the physical constraints, and motivates the design and",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
190,
325,
203
],
"spans": [
{
"bbox": [
105,
190,
325,
203
],
"score": 1.0,
"content": "development of more reliable and robust BCI systems.",
"type": "text"
}
],
"index": 9
}
],
"index": 5,
"bbox_fs": [
105,
101,
507,
203
]
},
{
"type": "title",
"bbox": [
108,
214,
184,
225
],
"lines": [
{
"bbox": [
106,
214,
185,
226
],
"spans": [
{
"bbox": [
106,
214,
185,
226
],
"score": 1.0,
"content": "REPRODUCIBLITY",
"type": "text"
}
],
"index": 10
}
],
"index": 10
},
{
"type": "text",
"bbox": [
108,
233,
504,
255
],
"lines": [
{
"bbox": [
104,
231,
506,
248
],
"spans": [
{
"bbox": [
104,
231,
506,
248
],
"score": 1.0,
"content": "A link to a anonymous downloadable source code of this work is submitted as supplementary",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
244,
149,
256
],
"spans": [
{
"bbox": [
106,
244,
149,
256
],
"score": 1.0,
"content": "materials.",
"type": "text"
}
],
"index": 12
}
],
"index": 11.5,
"bbox_fs": [
104,
231,
506,
256
]
},
{
"type": "title",
"bbox": [
108,
272,
175,
285
],
"lines": [
{
"bbox": [
106,
272,
176,
286
],
"spans": [
{
"bbox": [
106,
272,
176,
286
],
"score": 1.0,
"content": "REFERENCES",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
107,
292,
384,
303
],
"lines": [
{
"bbox": [
105,
290,
385,
306
],
"spans": [
{
"bbox": [
105,
290,
385,
306
],
"score": 1.0,
"content": "BCI2000: A general-purpose brain-computer interface (BCI) system.",
"type": "text"
}
],
"index": 14
}
],
"index": 14,
"bbox_fs": [
105,
290,
385,
306
]
},
{
"type": "text",
"bbox": [
104,
312,
505,
334
],
"lines": [
{
"bbox": [
105,
311,
506,
325
],
"spans": [
{
"bbox": [
105,
311,
506,
325
],
"score": 1.0,
"content": "Ayten Ozge Akmandor and Niraj K. Jha. Smart health care: An edge-side computing perspective.",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
115,
322,
349,
335
],
"spans": [
{
"bbox": [
115,
322,
349,
335
],
"score": 1.0,
"content": "IEEE Consumer Electronics Magazine, 7(1):29–37, 2018.",
"type": "text"
}
],
"index": 16
}
],
"index": 15.5,
"bbox_fs": [
105,
311,
506,
335
]
},
{
"type": "text",
"bbox": [
107,
342,
504,
366
],
"lines": [
{
"bbox": [
106,
342,
504,
355
],
"spans": [
{
"bbox": [
106,
342,
504,
355
],
"score": 1.0,
"content": "Ralph Algazi, Carlos Avendano, and Richard O Duda. Estimation of a spherical-head model from",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
116,
353,
293,
366
],
"spans": [
{
"bbox": [
116,
353,
293,
366
],
"score": 1.0,
"content": "anthropometry. J. Aud. Eng. Soc., 49, 2001.",
"type": "text"
}
],
"index": 18
}
],
"index": 17.5,
"bbox_fs": [
106,
342,
504,
366
]
},
{
"type": "text",
"bbox": [
106,
373,
503,
396
],
"lines": [
{
"bbox": [
105,
373,
505,
386
],
"spans": [
{
"bbox": [
105,
373,
505,
386
],
"score": 1.0,
"content": "Pietro Aricò, Nicolina Sciaraffa, and Fabio Babiloni. Brain–computer interfaces: Toward a daily life",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
115,
385,
288,
397
],
"spans": [
{
"bbox": [
115,
385,
288,
397
],
"score": 1.0,
"content": "employment. Brain Sciences, 10(3), 2020.",
"type": "text"
}
],
"index": 20
}
],
"index": 19.5,
"bbox_fs": [
105,
373,
505,
397
]
},
{
"type": "text",
"bbox": [
106,
404,
506,
439
],
"lines": [
{
"bbox": [
106,
405,
505,
417
],
"spans": [
{
"bbox": [
106,
405,
505,
417
],
"score": 1.0,
"content": "Aamir Arsalan, Muhammad Majid, Amna Rauf Butt, and Syed Muhammad Anwar. Classification",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
116,
415,
506,
429
],
"spans": [
{
"bbox": [
116,
415,
506,
429
],
"score": 1.0,
"content": "of perceived mental stress using a commercially available EEG headband. IEEE Journal of",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
116,
427,
361,
439
],
"spans": [
{
"bbox": [
116,
427,
361,
439
],
"score": 1.0,
"content": "Biomedical and Health Informatics, 23(6):2257–2264, 2019.",
"type": "text"
}
],
"index": 23
}
],
"index": 22,
"bbox_fs": [
106,
405,
506,
439
]
},
{
"type": "text",
"bbox": [
108,
446,
505,
481
],
"lines": [
{
"bbox": [
106,
446,
505,
459
],
"spans": [
{
"bbox": [
106,
446,
505,
459
],
"score": 1.0,
"content": "Christopher Beach, Ertan Balaban, and Alexander J. Casson. Chapter 14 - edge algorithms for",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
115,
458,
505,
470
],
"spans": [
{
"bbox": [
115,
458,
505,
470
],
"score": 1.0,
"content": "wearables: an overview of a truly multi-disciplinary problem. In Edward Sazonov (ed.), Wearable",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
115,
469,
497,
481
],
"spans": [
{
"bbox": [
115,
469,
497,
481
],
"score": 1.0,
"content": "Sensors (Second Edition), pp. 379–414. Academic Press, Oxford, second edition edition, 2021.",
"type": "text"
}
],
"index": 26
}
],
"index": 25,
"bbox_fs": [
106,
446,
505,
481
]
},
{
"type": "text",
"bbox": [
108,
488,
504,
523
],
"lines": [
{
"bbox": [
106,
488,
505,
501
],
"spans": [
{
"bbox": [
106,
488,
505,
501
],
"score": 1.0,
"content": "Kais Belwafi, Olivier Romain, Sofien Gannouni, Fakhreddine Ghaffari, Ridha Djemal, and Bouraoui",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
115,
500,
506,
513
],
"spans": [
{
"bbox": [
115,
500,
506,
513
],
"score": 1.0,
"content": "Ouni. An embedded implementation based on adaptive filter bank for brain–computer interface",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
115,
511,
318,
523
],
"spans": [
{
"bbox": [
115,
511,
318,
523
],
"score": 1.0,
"content": "systems. Journal of Neuroscience Methods, 2018.",
"type": "text"
}
],
"index": 29
}
],
"index": 28,
"bbox_fs": [
106,
488,
506,
523
]
},
{
"type": "text",
"bbox": [
107,
530,
506,
564
],
"lines": [
{
"bbox": [
105,
529,
506,
544
],
"spans": [
{
"bbox": [
105,
529,
506,
544
],
"score": 1.0,
"content": "Sergio López Bernal, Alberto Huertas Celdrán, Gregorio Martínez Pérez, Michael Taynnan Barros,",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
116,
542,
506,
554
],
"spans": [
{
"bbox": [
116,
542,
506,
554
],
"score": 1.0,
"content": "and Sasitharan Balasubramaniam. Security in brain-computer interfaces: State-of-the-art,",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
115,
552,
436,
565
],
"spans": [
{
"bbox": [
115,
552,
436,
565
],
"score": 1.0,
"content": "opportunities, and future challenges. ACM Comput. Surv., 54(1), January 2021.",
"type": "text"
}
],
"index": 32
}
],
"index": 31,
"bbox_fs": [
105,
529,
506,
565
]
},
{
"type": "text",
"bbox": [
106,
572,
506,
617
],
"lines": [
{
"bbox": [
105,
572,
506,
585
],
"spans": [
{
"bbox": [
105,
572,
506,
585
],
"score": 1.0,
"content": "Florian C. A. A. Bodranghien, Margot Langlois Mahe, Serge Clément, and Mario U. Manto. A pilot",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
115,
582,
506,
597
],
"spans": [
{
"bbox": [
115,
582,
506,
597
],
"score": 1.0,
"content": "study on the effects of transcranial direct current stimulation on brain rhythms and entropy during",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
115,
595,
507,
608
],
"spans": [
{
"bbox": [
115,
595,
507,
608
],
"score": 1.0,
"content": "self-paced finger movement using the Epoc helmet. Frontiers in Human Neuroscience, 11:201,",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
115,
604,
143,
617
],
"spans": [
{
"bbox": [
115,
604,
143,
617
],
"score": 1.0,
"content": "2017.",
"type": "text"
}
],
"index": 36
}
],
"index": 34.5,
"bbox_fs": [
105,
572,
507,
617
]
},
{
"type": "text",
"bbox": [
107,
625,
506,
670
],
"lines": [
{
"bbox": [
105,
626,
506,
639
],
"spans": [
{
"bbox": [
105,
626,
506,
639
],
"score": 1.0,
"content": "Jonathan S. Brumberg, Jeremy D. Burnison, and Kevin M. Pitt. Using motor imagery to control",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
115,
636,
506,
649
],
"spans": [
{
"bbox": [
115,
636,
506,
649
],
"score": 1.0,
"content": "brain-computer interfaces for communication. In Dylan D. Schmorrow and Cali M. Fidopiastis",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
115,
647,
507,
662
],
"spans": [
{
"bbox": [
115,
647,
507,
662
],
"score": 1.0,
"content": "(eds.), Foundations of Augmented Cognition: Neuroergonomics and Operational Neuroscience, pp.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
115,
658,
337,
672
],
"spans": [
{
"bbox": [
115,
658,
337,
672
],
"score": 1.0,
"content": "14–25, Cham, 2016. Springer International Publishing.",
"type": "text"
}
],
"index": 40
}
],
"index": 38.5,
"bbox_fs": [
105,
626,
507,
672
]
},
{
"type": "text",
"bbox": [
106,
678,
505,
701
],
"lines": [
{
"bbox": [
105,
678,
505,
691
],
"spans": [
{
"bbox": [
105,
678,
505,
691
],
"score": 1.0,
"content": "C Brunner, R Leeb, G R Müller-Putz, A Schlögl, and G Pfurtscheller. BCI competition 2008 - Graz",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
115,
688,
492,
702
],
"spans": [
{
"bbox": [
115,
688,
492,
702
],
"score": 1.0,
"content": "data set A, 2008. URL http://bnci-horizon-2020.eu/database/data-sets.",
"type": "text"
}
],
"index": 42
}
],
"index": 41.5,
"bbox_fs": [
105,
678,
505,
702
]
},
{
"type": "text",
"bbox": [
108,
709,
504,
732
],
"lines": [
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "Ujwal Chaudhary, Niels Birbaumer, and Ander Ramos-Murguialday. Brain–computer interfaces for",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
116,
720,
420,
733
],
"spans": [
{
"bbox": [
116,
720,
420,
733
],
"score": 1.0,
"content": "communication and rehabilitation. Nature Reviews Neurology, 12, 08 2016.",
"type": "text"
}
],
"index": 44
}
],
"index": 43.5,
"bbox_fs": [
106,
709,
505,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
108,
82,
504,
116
],
"lines": [
{
"bbox": [
104,
80,
505,
97
],
"spans": [
{
"bbox": [
104,
80,
505,
97
],
"score": 1.0,
"content": "Hauke Dose, Jakob S. Moller, Sadasivan Puthusserypady, and Helle K. Iversen. A deep learning",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
115,
92,
505,
107
],
"spans": [
{
"bbox": [
115,
92,
505,
107
],
"score": 1.0,
"content": "MI - EEG classification model for BCIs. In 2018 26th European Signal Processing Conference",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
115,
104,
285,
117
],
"spans": [
{
"bbox": [
115,
104,
285,
117
],
"score": 1.0,
"content": "(EUSIPCO), pp. 1676–1679. IEEE, 2018.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "text",
"bbox": [
105,
122,
504,
145
],
"lines": [
{
"bbox": [
106,
122,
506,
135
],
"spans": [
{
"bbox": [
106,
122,
506,
135
],
"score": 1.0,
"content": "Anirban Dutta. Brain–computer interface spellers for communication: Why we need to address their",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
115,
134,
354,
146
],
"spans": [
{
"bbox": [
115,
134,
354,
146
],
"score": 1.0,
"content": "security and authenticity. Brain Sciences, 10:139, 03 2020.",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5
},
{
"type": "text",
"bbox": [
106,
151,
504,
175
],
"lines": [
{
"bbox": [
105,
149,
507,
166
],
"spans": [
{
"bbox": [
105,
149,
507,
166
],
"score": 1.0,
"content": "Daniel Elstob and Emanuele Secco. A low cost EEG based BCI prosthetic using motor imagery.",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
115,
162,
498,
175
],
"spans": [
{
"bbox": [
115,
162,
498,
175
],
"score": 1.0,
"content": "International Journal of Information Technology Convergence and Services, 6:23–36, 02 2016.",
"type": "text"
}
],
"index": 6
}
],
"index": 5.5
},
{
"type": "text",
"bbox": [
106,
180,
504,
214
],
"lines": [
{
"bbox": [
106,
180,
505,
193
],
"spans": [
{
"bbox": [
106,
180,
505,
193
],
"score": 1.0,
"content": "Anna Fertonani, Clarissa Ferrari, and Carlo Miniussi. What do you feel if I apply transcranial electric",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
115,
191,
506,
205
],
"spans": [
{
"bbox": [
115,
191,
506,
205
],
"score": 1.0,
"content": "stimulation? Safety, sensations and secondary induced effects. Clinical Neurophysiology, 126(11):",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
115,
203,
191,
214
],
"spans": [
{
"bbox": [
115,
203,
191,
214
],
"score": 1.0,
"content": "2181–2188, 2015.",
"type": "text"
}
],
"index": 9
}
],
"index": 8
},
{
"type": "text",
"bbox": [
106,
220,
505,
244
],
"lines": [
{
"bbox": [
106,
221,
506,
233
],
"spans": [
{
"bbox": [
106,
221,
506,
233
],
"score": 1.0,
"content": "Samuel G. Finlayson, John D. Bowers, Joichi Ito, Jonathan L. Zittrain, Andrew L. Beam, and Isaac S.",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
116,
232,
506,
244
],
"spans": [
{
"bbox": [
116,
232,
506,
244
],
"score": 1.0,
"content": "Kohane. Adversarial attacks on medical machine learning. Science, 363(6433):1287–1289, 2019.",
"type": "text"
}
],
"index": 11
}
],
"index": 10.5
},
{
"type": "text",
"bbox": [
106,
250,
505,
273
],
"lines": [
{
"bbox": [
106,
250,
506,
263
],
"spans": [
{
"bbox": [
106,
250,
506,
263
],
"score": 1.0,
"content": "Flowneuroscience. Flow tDCS device. https://flowneuroscience.com/, 2021. Accessed:",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
116,
260,
169,
272
],
"spans": [
{
"bbox": [
116,
260,
169,
272
],
"score": 1.0,
"content": "2021-05-27.",
"type": "text"
}
],
"index": 13
}
],
"index": 12.5
},
{
"type": "text",
"bbox": [
104,
279,
505,
302
],
"lines": [
{
"bbox": [
106,
279,
505,
293
],
"spans": [
{
"bbox": [
106,
279,
505,
293
],
"score": 1.0,
"content": "Daniel Freer and Guang-Zhong Yang. Data augmentation for self-paced motor imagery classification",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
115,
290,
359,
303
],
"spans": [
{
"bbox": [
115,
290,
359,
303
],
"score": 1.0,
"content": "with C-LSTM. Journal of Neural Engineering, 17, 01 2020.",
"type": "text"
}
],
"index": 15
}
],
"index": 14.5
},
{
"type": "text",
"bbox": [
107,
307,
505,
353
],
"lines": [
{
"bbox": [
106,
309,
505,
321
],
"spans": [
{
"bbox": [
106,
309,
505,
321
],
"score": 1.0,
"content": "Ary L Goldberger, Luis A N Amaral, Leon Glass, Jeffrey M Hausdorff, Plamen Ch Ivanov, Roger G",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
114,
318,
507,
332
],
"spans": [
{
"bbox": [
114,
318,
507,
332
],
"score": 1.0,
"content": "Mark, Joseph E Mietus, George B Moody, Chung-Kang Peng, and H Eugene Stanley. PhysioBank,",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
115,
330,
505,
344
],
"spans": [
{
"bbox": [
115,
330,
505,
344
],
"score": 1.0,
"content": "PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
115,
341,
307,
354
],
"spans": [
{
"bbox": [
115,
341,
307,
354
],
"score": 1.0,
"content": "signals. circulation, 101(23):e215–e220, 2000.",
"type": "text"
}
],
"index": 19
}
],
"index": 17.5
},
{
"type": "text",
"bbox": [
104,
359,
504,
383
],
"lines": [
{
"bbox": [
105,
358,
505,
372
],
"spans": [
{
"bbox": [
105,
358,
505,
372
],
"score": 1.0,
"content": "Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
115,
370,
470,
383
],
"spans": [
{
"bbox": [
115,
370,
470,
383
],
"score": 1.0,
"content": "examples. In Proc. International Conference on Learning Representations, ICLR, 2015.",
"type": "text"
}
],
"index": 21
}
],
"index": 20.5
},
{
"type": "text",
"bbox": [
106,
388,
505,
422
],
"lines": [
{
"bbox": [
105,
387,
505,
402
],
"spans": [
{
"bbox": [
105,
387,
505,
402
],
"score": 1.0,
"content": "Xintian Han, Yuxuan Hu, Luca Foschini, Larry Chinitz, Lior Jankelson, and Rajesh Ranganath. Deep",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
116,
399,
505,
411
],
"spans": [
{
"bbox": [
116,
399,
505,
411
],
"score": 1.0,
"content": "learning models for electrocardiograms are susceptible to adversarial attack. Nature medicine, 26",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
115,
410,
195,
422
],
"spans": [
{
"bbox": [
115,
410,
195,
422
],
"score": 1.0,
"content": "(3):360–363, 2020.",
"type": "text"
}
],
"index": 24
}
],
"index": 23
},
{
"type": "text",
"bbox": [
106,
428,
504,
462
],
"lines": [
{
"bbox": [
106,
428,
506,
442
],
"spans": [
{
"bbox": [
106,
428,
506,
442
],
"score": 1.0,
"content": "Xue Jiang, Xiao Zhang, and Dongrui Wu. Active learning for black-box adversarial attacks in EEG-",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
116,
440,
505,
452
],
"spans": [
{
"bbox": [
116,
440,
505,
452
],
"score": 1.0,
"content": "based brain-computer interfaces. In 2019 IEEE Symposium Series on Computational Intelligence",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
116,
451,
227,
463
],
"spans": [
{
"bbox": [
116,
451,
227,
463
],
"score": 1.0,
"content": "(SSCI), pp. 361–368, 2019.",
"type": "text"
}
],
"index": 27
}
],
"index": 26
},
{
"type": "text",
"bbox": [
104,
468,
505,
492
],
"lines": [
{
"bbox": [
106,
468,
505,
481
],
"spans": [
{
"bbox": [
106,
468,
505,
481
],
"score": 1.0,
"content": "Mandar Karlekar and Anubha Gupta. Stochastic modeling of EEG rhythms with fractional gaussian",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
115,
480,
506,
492
],
"spans": [
{
"bbox": [
115,
480,
506,
492
],
"score": 1.0,
"content": "noise. In 2014 22nd European Signal Processing Conference (EUSIPCO), pp. 2520–2524, 2014.",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5
},
{
"type": "text",
"bbox": [
106,
497,
505,
542
],
"lines": [
{
"bbox": [
106,
498,
505,
510
],
"spans": [
{
"bbox": [
106,
498,
505,
510
],
"score": 1.0,
"content": "Victor Kartsch, Giuseppe Tagliavini, Marco Guermandi, Simone Benatti, Davide Rossi, and Luca",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
115,
509,
506,
522
],
"spans": [
{
"bbox": [
115,
509,
506,
522
],
"score": 1.0,
"content": "Benini. Biowolf: A sub-10-mW 8-channel advanced brain–computer interface platform with a",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
115,
520,
506,
532
],
"spans": [
{
"bbox": [
115,
520,
506,
532
],
"score": 1.0,
"content": "nine-core processor and BLE connectivity. IEEE Transactions on Biomedical Circuits and Systems,",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
115,
531,
205,
543
],
"spans": [
{
"bbox": [
115,
531,
205,
543
],
"score": 1.0,
"content": "13(5):893–906, 2019.",
"type": "text"
}
],
"index": 33
}
],
"index": 31.5
},
{
"type": "text",
"bbox": [
106,
548,
505,
583
],
"lines": [
{
"bbox": [
105,
548,
506,
562
],
"spans": [
{
"bbox": [
105,
548,
506,
562
],
"score": 1.0,
"content": "Min Kim, Youngchang Cho, Suk-Tae Seo, Chang-Sik Son, Hee-Joon Park, and Yoon-Nyun Kim.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
116,
560,
504,
573
],
"spans": [
{
"bbox": [
116,
560,
504,
573
],
"score": 1.0,
"content": "A new method for non-invasive measurement of skin in the low frequency range. Healthcare",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
116,
571,
283,
583
],
"spans": [
{
"bbox": [
116,
571,
283,
583
],
"score": 1.0,
"content": "informatics research, 16:143–8, 09 2010.",
"type": "text"
}
],
"index": 36
}
],
"index": 35
},
{
"type": "text",
"bbox": [
105,
588,
505,
612
],
"lines": [
{
"bbox": [
106,
588,
506,
601
],
"spans": [
{
"bbox": [
106,
588,
506,
601
],
"score": 1.0,
"content": "Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world.",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
116,
600,
213,
611
],
"spans": [
{
"bbox": [
116,
600,
213,
611
],
"score": 1.0,
"content": "CoRR, abs/1607.02533.",
"type": "text"
}
],
"index": 38
}
],
"index": 37.5
},
{
"type": "text",
"bbox": [
107,
618,
505,
653
],
"lines": [
{
"bbox": [
106,
618,
505,
631
],
"spans": [
{
"bbox": [
106,
618,
505,
631
],
"score": 1.0,
"content": "Vernon J Lawhern, Amelia J Solon, Nicholas R Waytowich, Stephen M Gordon, Chou P Hung, and",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
116,
629,
505,
642
],
"spans": [
{
"bbox": [
116,
629,
505,
642
],
"score": 1.0,
"content": "Brent J Lance. EEGNet: a compact convolutional neural network for EEG-based brain–computer",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
116,
640,
381,
653
],
"spans": [
{
"bbox": [
116,
640,
381,
653
],
"score": 1.0,
"content": "interfaces. Journal of Neural Engineering, 15(5):056013, 7 2018.",
"type": "text"
}
],
"index": 41
}
],
"index": 40
},
{
"type": "text",
"bbox": [
106,
658,
504,
692
],
"lines": [
{
"bbox": [
106,
658,
505,
671
],
"spans": [
{
"bbox": [
106,
658,
505,
671
],
"score": 1.0,
"content": "Jinfeng Li, Shouling Ji, Tianyu Du, Bo Li, and Ting Wang. Textbugger: Generating adversarial text",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
115,
669,
506,
683
],
"spans": [
{
"bbox": [
115,
669,
506,
683
],
"score": 1.0,
"content": "against real-world applications. In Proc. Network and Distributed System Security Symposium,",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
115,
681,
260,
693
],
"spans": [
{
"bbox": [
115,
681,
260,
693
],
"score": 1.0,
"content": "NDSS. The Internet Society, 2019a.",
"type": "text"
}
],
"index": 44
}
],
"index": 43
},
{
"type": "text",
"bbox": [
107,
698,
505,
732
],
"lines": [
{
"bbox": [
106,
699,
505,
711
],
"spans": [
{
"bbox": [
106,
699,
505,
711
],
"score": 1.0,
"content": "Juncheng Li, Shuhui Qu, Xinjian Li, Joseph Szurley, J. Zico Kolter, and Florian Metze. Adversarial",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
117,
710,
504,
721
],
"spans": [
{
"bbox": [
117,
710,
504,
721
],
"score": 1.0,
"content": "music: Real world audio adversary against wake-word detection system. In Proc. Neural",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
115,
720,
398,
734
],
"spans": [
{
"bbox": [
115,
720,
398,
734
],
"score": 1.0,
"content": "Information Processing Systems (NeurIPS), pp. 11908–11918, 2019b.",
"type": "text"
}
],
"index": 47
}
],
"index": 46
}
],
"page_idx": 10,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
300,
751,
310,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
765
],
"spans": [
{
"bbox": [
299,
750,
312,
765
],
"score": 1.0,
"content": "",
"type": "text",
"height": 15,
"width": 13
}
]
}
]
},
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
108,
82,
504,
116
],
"lines": [
{
"bbox": [
104,
80,
505,
97
],
"spans": [
{
"bbox": [
104,
80,
505,
97
],
"score": 1.0,
"content": "Hauke Dose, Jakob S. Moller, Sadasivan Puthusserypady, and Helle K. Iversen. A deep learning",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
115,
92,
505,
107
],
"spans": [
{
"bbox": [
115,
92,
505,
107
],
"score": 1.0,
"content": "MI - EEG classification model for BCIs. In 2018 26th European Signal Processing Conference",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
115,
104,
285,
117
],
"spans": [
{
"bbox": [
115,
104,
285,
117
],
"score": 1.0,
"content": "(EUSIPCO), pp. 1676–1679. IEEE, 2018.",
"type": "text"
}
],
"index": 2
}
],
"index": 1,
"bbox_fs": [
104,
80,
505,
117
]
},
{
"type": "text",
"bbox": [
105,
122,
504,
145
],
"lines": [
{
"bbox": [
106,
122,
506,
135
],
"spans": [
{
"bbox": [
106,
122,
506,
135
],
"score": 1.0,
"content": "Anirban Dutta. Brain–computer interface spellers for communication: Why we need to address their",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
115,
134,
354,
146
],
"spans": [
{
"bbox": [
115,
134,
354,
146
],
"score": 1.0,
"content": "security and authenticity. Brain Sciences, 10:139, 03 2020.",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5,
"bbox_fs": [
106,
122,
506,
146
]
},
{
"type": "text",
"bbox": [
106,
151,
504,
175
],
"lines": [
{
"bbox": [
105,
149,
507,
166
],
"spans": [
{
"bbox": [
105,
149,
507,
166
],
"score": 1.0,
"content": "Daniel Elstob and Emanuele Secco. A low cost EEG based BCI prosthetic using motor imagery.",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
115,
162,
498,
175
],
"spans": [
{
"bbox": [
115,
162,
498,
175
],
"score": 1.0,
"content": "International Journal of Information Technology Convergence and Services, 6:23–36, 02 2016.",
"type": "text"
}
],
"index": 6
}
],
"index": 5.5,
"bbox_fs": [
105,
149,
507,
175
]
},
{
"type": "text",
"bbox": [
106,
180,
504,
214
],
"lines": [
{
"bbox": [
106,
180,
505,
193
],
"spans": [
{
"bbox": [
106,
180,
505,
193
],
"score": 1.0,
"content": "Anna Fertonani, Clarissa Ferrari, and Carlo Miniussi. What do you feel if I apply transcranial electric",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
115,
191,
506,
205
],
"spans": [
{
"bbox": [
115,
191,
506,
205
],
"score": 1.0,
"content": "stimulation? Safety, sensations and secondary induced effects. Clinical Neurophysiology, 126(11):",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
115,
203,
191,
214
],
"spans": [
{
"bbox": [
115,
203,
191,
214
],
"score": 1.0,
"content": "2181–2188, 2015.",
"type": "text"
}
],
"index": 9
}
],
"index": 8,
"bbox_fs": [
106,
180,
506,
214
]
},
{
"type": "text",
"bbox": [
106,
220,
505,
244
],
"lines": [
{
"bbox": [
106,
221,
506,
233
],
"spans": [
{
"bbox": [
106,
221,
506,
233
],
"score": 1.0,
"content": "Samuel G. Finlayson, John D. Bowers, Joichi Ito, Jonathan L. Zittrain, Andrew L. Beam, and Isaac S.",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
116,
232,
506,
244
],
"spans": [
{
"bbox": [
116,
232,
506,
244
],
"score": 1.0,
"content": "Kohane. Adversarial attacks on medical machine learning. Science, 363(6433):1287–1289, 2019.",
"type": "text"
}
],
"index": 11
}
],
"index": 10.5,
"bbox_fs": [
106,
221,
506,
244
]
},
{
"type": "text",
"bbox": [
106,
250,
505,
273
],
"lines": [
{
"bbox": [
106,
250,
506,
263
],
"spans": [
{
"bbox": [
106,
250,
506,
263
],
"score": 1.0,
"content": "Flowneuroscience. Flow tDCS device. https://flowneuroscience.com/, 2021. Accessed:",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
116,
260,
169,
272
],
"spans": [
{
"bbox": [
116,
260,
169,
272
],
"score": 1.0,
"content": "2021-05-27.",
"type": "text"
}
],
"index": 13
}
],
"index": 12.5,
"bbox_fs": [
106,
250,
506,
272
]
},
{
"type": "text",
"bbox": [
104,
279,
505,
302
],
"lines": [
{
"bbox": [
106,
279,
505,
293
],
"spans": [
{
"bbox": [
106,
279,
505,
293
],
"score": 1.0,
"content": "Daniel Freer and Guang-Zhong Yang. Data augmentation for self-paced motor imagery classification",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
115,
290,
359,
303
],
"spans": [
{
"bbox": [
115,
290,
359,
303
],
"score": 1.0,
"content": "with C-LSTM. Journal of Neural Engineering, 17, 01 2020.",
"type": "text"
}
],
"index": 15
}
],
"index": 14.5,
"bbox_fs": [
106,
279,
505,
303
]
},
{
"type": "text",
"bbox": [
107,
307,
505,
353
],
"lines": [
{
"bbox": [
106,
309,
505,
321
],
"spans": [
{
"bbox": [
106,
309,
505,
321
],
"score": 1.0,
"content": "Ary L Goldberger, Luis A N Amaral, Leon Glass, Jeffrey M Hausdorff, Plamen Ch Ivanov, Roger G",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
114,
318,
507,
332
],
"spans": [
{
"bbox": [
114,
318,
507,
332
],
"score": 1.0,
"content": "Mark, Joseph E Mietus, George B Moody, Chung-Kang Peng, and H Eugene Stanley. PhysioBank,",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
115,
330,
505,
344
],
"spans": [
{
"bbox": [
115,
330,
505,
344
],
"score": 1.0,
"content": "PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
115,
341,
307,
354
],
"spans": [
{
"bbox": [
115,
341,
307,
354
],
"score": 1.0,
"content": "signals. circulation, 101(23):e215–e220, 2000.",
"type": "text"
}
],
"index": 19
}
],
"index": 17.5,
"bbox_fs": [
106,
309,
507,
354
]
},
{
"type": "text",
"bbox": [
104,
359,
504,
383
],
"lines": [
{
"bbox": [
105,
358,
505,
372
],
"spans": [
{
"bbox": [
105,
358,
505,
372
],
"score": 1.0,
"content": "Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
115,
370,
470,
383
],
"spans": [
{
"bbox": [
115,
370,
470,
383
],
"score": 1.0,
"content": "examples. In Proc. International Conference on Learning Representations, ICLR, 2015.",
"type": "text"
}
],
"index": 21
}
],
"index": 20.5,
"bbox_fs": [
105,
358,
505,
383
]
},
{
"type": "text",
"bbox": [
106,
388,
505,
422
],
"lines": [
{
"bbox": [
105,
387,
505,
402
],
"spans": [
{
"bbox": [
105,
387,
505,
402
],
"score": 1.0,
"content": "Xintian Han, Yuxuan Hu, Luca Foschini, Larry Chinitz, Lior Jankelson, and Rajesh Ranganath. Deep",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
116,
399,
505,
411
],
"spans": [
{
"bbox": [
116,
399,
505,
411
],
"score": 1.0,
"content": "learning models for electrocardiograms are susceptible to adversarial attack. Nature medicine, 26",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
115,
410,
195,
422
],
"spans": [
{
"bbox": [
115,
410,
195,
422
],
"score": 1.0,
"content": "(3):360–363, 2020.",
"type": "text"
}
],
"index": 24
}
],
"index": 23,
"bbox_fs": [
105,
387,
505,
422
]
},
{
"type": "text",
"bbox": [
106,
428,
504,
462
],
"lines": [
{
"bbox": [
106,
428,
506,
442
],
"spans": [
{
"bbox": [
106,
428,
506,
442
],
"score": 1.0,
"content": "Xue Jiang, Xiao Zhang, and Dongrui Wu. Active learning for black-box adversarial attacks in EEG-",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
116,
440,
505,
452
],
"spans": [
{
"bbox": [
116,
440,
505,
452
],
"score": 1.0,
"content": "based brain-computer interfaces. In 2019 IEEE Symposium Series on Computational Intelligence",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
116,
451,
227,
463
],
"spans": [
{
"bbox": [
116,
451,
227,
463
],
"score": 1.0,
"content": "(SSCI), pp. 361–368, 2019.",
"type": "text"
}
],
"index": 27
}
],
"index": 26,
"bbox_fs": [
106,
428,
506,
463
]
},
{
"type": "text",
"bbox": [
104,
468,
505,
492
],
"lines": [
{
"bbox": [
106,
468,
505,
481
],
"spans": [
{
"bbox": [
106,
468,
505,
481
],
"score": 1.0,
"content": "Mandar Karlekar and Anubha Gupta. Stochastic modeling of EEG rhythms with fractional gaussian",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
115,
480,
506,
492
],
"spans": [
{
"bbox": [
115,
480,
506,
492
],
"score": 1.0,
"content": "noise. In 2014 22nd European Signal Processing Conference (EUSIPCO), pp. 2520–2524, 2014.",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5,
"bbox_fs": [
106,
468,
506,
492
]
},
{
"type": "text",
"bbox": [
106,
497,
505,
542
],
"lines": [
{
"bbox": [
106,
498,
505,
510
],
"spans": [
{
"bbox": [
106,
498,
505,
510
],
"score": 1.0,
"content": "Victor Kartsch, Giuseppe Tagliavini, Marco Guermandi, Simone Benatti, Davide Rossi, and Luca",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
115,
509,
506,
522
],
"spans": [
{
"bbox": [
115,
509,
506,
522
],
"score": 1.0,
"content": "Benini. Biowolf: A sub-10-mW 8-channel advanced brain–computer interface platform with a",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
115,
520,
506,
532
],
"spans": [
{
"bbox": [
115,
520,
506,
532
],
"score": 1.0,
"content": "nine-core processor and BLE connectivity. IEEE Transactions on Biomedical Circuits and Systems,",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
115,
531,
205,
543
],
"spans": [
{
"bbox": [
115,
531,
205,
543
],
"score": 1.0,
"content": "13(5):893–906, 2019.",
"type": "text"
}
],
"index": 33
}
],
"index": 31.5,
"bbox_fs": [
106,
498,
506,
543
]
},
{
"type": "text",
"bbox": [
106,
548,
505,
583
],
"lines": [
{
"bbox": [
105,
548,
506,
562
],
"spans": [
{
"bbox": [
105,
548,
506,
562
],
"score": 1.0,
"content": "Min Kim, Youngchang Cho, Suk-Tae Seo, Chang-Sik Son, Hee-Joon Park, and Yoon-Nyun Kim.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
116,
560,
504,
573
],
"spans": [
{
"bbox": [
116,
560,
504,
573
],
"score": 1.0,
"content": "A new method for non-invasive measurement of skin in the low frequency range. Healthcare",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
116,
571,
283,
583
],
"spans": [
{
"bbox": [
116,
571,
283,
583
],
"score": 1.0,
"content": "informatics research, 16:143–8, 09 2010.",
"type": "text"
}
],
"index": 36
}
],
"index": 35,
"bbox_fs": [
105,
548,
506,
583
]
},
{
"type": "text",
"bbox": [
105,
588,
505,
612
],
"lines": [
{
"bbox": [
106,
588,
506,
601
],
"spans": [
{
"bbox": [
106,
588,
506,
601
],
"score": 1.0,
"content": "Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world.",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
116,
600,
213,
611
],
"spans": [
{
"bbox": [
116,
600,
213,
611
],
"score": 1.0,
"content": "CoRR, abs/1607.02533.",
"type": "text"
}
],
"index": 38
}
],
"index": 37.5,
"bbox_fs": [
106,
588,
506,
611
]
},
{
"type": "text",
"bbox": [
107,
618,
505,
653
],
"lines": [
{
"bbox": [
106,
618,
505,
631
],
"spans": [
{
"bbox": [
106,
618,
505,
631
],
"score": 1.0,
"content": "Vernon J Lawhern, Amelia J Solon, Nicholas R Waytowich, Stephen M Gordon, Chou P Hung, and",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
116,
629,
505,
642
],
"spans": [
{
"bbox": [
116,
629,
505,
642
],
"score": 1.0,
"content": "Brent J Lance. EEGNet: a compact convolutional neural network for EEG-based brain–computer",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
116,
640,
381,
653
],
"spans": [
{
"bbox": [
116,
640,
381,
653
],
"score": 1.0,
"content": "interfaces. Journal of Neural Engineering, 15(5):056013, 7 2018.",
"type": "text"
}
],
"index": 41
}
],
"index": 40,
"bbox_fs": [
106,
618,
505,
653
]
},
{
"type": "text",
"bbox": [
106,
658,
504,
692
],
"lines": [
{
"bbox": [
106,
658,
505,
671
],
"spans": [
{
"bbox": [
106,
658,
505,
671
],
"score": 1.0,
"content": "Jinfeng Li, Shouling Ji, Tianyu Du, Bo Li, and Ting Wang. Textbugger: Generating adversarial text",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
115,
669,
506,
683
],
"spans": [
{
"bbox": [
115,
669,
506,
683
],
"score": 1.0,
"content": "against real-world applications. In Proc. Network and Distributed System Security Symposium,",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
115,
681,
260,
693
],
"spans": [
{
"bbox": [
115,
681,
260,
693
],
"score": 1.0,
"content": "NDSS. The Internet Society, 2019a.",
"type": "text"
}
],
"index": 44
}
],
"index": 43,
"bbox_fs": [
106,
658,
506,
693
]
},
{
"type": "text",
"bbox": [
107,
698,
505,
732
],
"lines": [
{
"bbox": [
106,
699,
505,
711
],
"spans": [
{
"bbox": [
106,
699,
505,
711
],
"score": 1.0,
"content": "Juncheng Li, Shuhui Qu, Xinjian Li, Joseph Szurley, J. Zico Kolter, and Florian Metze. Adversarial",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
117,
710,
504,
721
],
"spans": [
{
"bbox": [
117,
710,
504,
721
],
"score": 1.0,
"content": "music: Real world audio adversary against wake-word detection system. In Proc. Neural",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
115,
720,
398,
734
],
"spans": [
{
"bbox": [
115,
720,
398,
734
],
"score": 1.0,
"content": "Information Processing Systems (NeurIPS), pp. 11908–11918, 2019b.",
"type": "text"
}
],
"index": 47
}
],
"index": 46,
"bbox_fs": [
106,
699,
505,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
105,
82,
504,
105
],
"lines": [
{
"bbox": [
106,
82,
505,
95
],
"spans": [
{
"bbox": [
106,
82,
505,
95
],
"score": 1.0,
"content": "Zihan Liu, Lubin Meng, Xiao Zhang, Weili Fang, and Dongrui Wu. Universal adversarial",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
115,
93,
360,
106
],
"spans": [
{
"bbox": [
115,
93,
360,
106
],
"score": 1.0,
"content": "perturbations for CNN classifiers in EEG-based BCIs, 2021.",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
108,
113,
504,
147
],
"lines": [
{
"bbox": [
106,
113,
505,
126
],
"spans": [
{
"bbox": [
106,
113,
505,
126
],
"score": 1.0,
"content": "Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu.",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
116,
124,
505,
137
],
"spans": [
{
"bbox": [
116,
124,
505,
137
],
"score": 1.0,
"content": "Towards deep learning models resistant to adversarial attacks. In Proc. International Conference",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
115,
136,
293,
147
],
"spans": [
{
"bbox": [
115,
136,
293,
147
],
"score": 1.0,
"content": "on Learning Representations (ICLR), 2018.",
"type": "text"
}
],
"index": 4
}
],
"index": 3
},
{
"type": "text",
"bbox": [
107,
155,
506,
189
],
"lines": [
{
"bbox": [
106,
155,
505,
167
],
"spans": [
{
"bbox": [
106,
155,
505,
167
],
"score": 1.0,
"content": "Alireza Malekmohammadi, Hoda Mohammadzade, Alireza Chamanzar, Mahdi Shabany, and",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
115,
166,
505,
178
],
"spans": [
{
"bbox": [
115,
166,
505,
178
],
"score": 1.0,
"content": "Benyamin Ghojogh. An efficient hardware implementation for a motor imagery brain computer",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
115,
177,
324,
189
],
"spans": [
{
"bbox": [
115,
177,
324,
189
],
"score": 1.0,
"content": "interface system. Scientia Iranica, 26:72–94, 2019.",
"type": "text"
}
],
"index": 7
}
],
"index": 6
},
{
"type": "text",
"bbox": [
106,
196,
506,
231
],
"lines": [
{
"bbox": [
105,
196,
505,
210
],
"spans": [
{
"bbox": [
105,
196,
505,
210
],
"score": 1.0,
"content": "Iuliana Marin, Myssar Jabbar Hammood Al-Battbootti, and Nicolae Goga. Drone control based on",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
115,
207,
506,
221
],
"spans": [
{
"bbox": [
115,
207,
506,
221
],
"score": 1.0,
"content": "mental commands and facial expressions. In 2020 12th International Conference on Electronics,",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
116,
218,
360,
232
],
"spans": [
{
"bbox": [
116,
218,
360,
232
],
"score": 1.0,
"content": "Computers and Artificial Intelligence (ECAI), pp. 1–4, 2020.",
"type": "text"
}
],
"index": 10
}
],
"index": 9
},
{
"type": "text",
"bbox": [
107,
238,
505,
272
],
"lines": [
{
"bbox": [
105,
237,
506,
252
],
"spans": [
{
"bbox": [
105,
237,
506,
252
],
"score": 1.0,
"content": "Lubin Meng, Jian Huang, Zhigang Zeng, Xue Jiang, Shan Yu, Tzyy-Ping Jung, Chin-Teng Lin,",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
116,
250,
505,
262
],
"spans": [
{
"bbox": [
116,
250,
505,
262
],
"score": 1.0,
"content": "Ricardo Chavarriaga, and Dongrui Wu. EEG-based brain-computer interfaces are vulnerable to",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
117,
261,
214,
273
],
"spans": [
{
"bbox": [
117,
261,
214,
273
],
"score": 1.0,
"content": "backdoor attacks, 2021.",
"type": "text"
}
],
"index": 13
}
],
"index": 12
},
{
"type": "text",
"bbox": [
107,
280,
506,
325
],
"lines": [
{
"bbox": [
105,
279,
507,
293
],
"spans": [
{
"bbox": [
105,
279,
507,
293
],
"score": 1.0,
"content": "Isabelle Merlet, Gwénaël Birot, Ricardo Salvador, Behnam Molaee-Ardekani, Abeye Mekonnen,",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
115,
290,
506,
305
],
"spans": [
{
"bbox": [
115,
290,
506,
305
],
"score": 1.0,
"content": "Aureli Soria-Frish, Giulio Ruffini, Pedro C. Miranda, and Fabrice Wendling. From oscillatory",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
115,
302,
506,
316
],
"spans": [
{
"bbox": [
115,
302,
506,
316
],
"score": 1.0,
"content": "transcranial current stimulation to scalp EEG changes: A biophysical and physiological modeling",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
117,
314,
265,
326
],
"spans": [
{
"bbox": [
117,
314,
265,
326
],
"score": 1.0,
"content": "study. PLOS ONE, 8:1–12, 02 2013.",
"type": "text"
}
],
"index": 17
}
],
"index": 15.5
},
{
"type": "text",
"bbox": [
107,
333,
505,
367
],
"lines": [
{
"bbox": [
106,
333,
505,
345
],
"spans": [
{
"bbox": [
106,
333,
505,
345
],
"score": 1.0,
"content": "Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
115,
344,
505,
357
],
"spans": [
{
"bbox": [
115,
344,
505,
357
],
"score": 1.0,
"content": "adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
115,
355,
248,
369
],
"spans": [
{
"bbox": [
115,
355,
248,
369
],
"score": 1.0,
"content": "Recognition (CVPR), July 2017.",
"type": "text"
}
],
"index": 20
}
],
"index": 19
},
{
"type": "text",
"bbox": [
107,
374,
504,
398
],
"lines": [
{
"bbox": [
106,
375,
505,
387
],
"spans": [
{
"bbox": [
106,
375,
505,
387
],
"score": 1.0,
"content": "G. Pfurtscheller and F.H. Lopes da Silva. Event-related EEG/MEG synchronization and",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
116,
385,
492,
398
],
"spans": [
{
"bbox": [
116,
385,
492,
398
],
"score": 1.0,
"content": "desynchronization: basic principles. Clinical Neurophysiology, 110(11):1842–1857, 11 1999.",
"type": "text"
}
],
"index": 22
}
],
"index": 21.5
},
{
"type": "text",
"bbox": [
106,
406,
505,
428
],
"lines": [
{
"bbox": [
105,
405,
505,
419
],
"spans": [
{
"bbox": [
105,
405,
505,
419
],
"score": 1.0,
"content": "Robert Plutchik and Henry R Hirsch. Skin impedance and phase angle as a function of frequency and",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
115,
417,
295,
428
],
"spans": [
{
"bbox": [
115,
417,
295,
428
],
"score": 1.0,
"content": "current. Science, 141(3584):927–928, 1963.",
"type": "text"
}
],
"index": 24
}
],
"index": 23.5
},
{
"type": "text",
"bbox": [
106,
436,
505,
471
],
"lines": [
{
"bbox": [
105,
435,
505,
450
],
"spans": [
{
"bbox": [
105,
435,
505,
450
],
"score": 1.0,
"content": "ZG Qiao, L Mørkrid, and NK Vøllestad. Characteristics of phase angle, extrapolated conductance",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
117,
449,
505,
460
],
"spans": [
{
"bbox": [
117,
449,
505,
460
],
"score": 1.0,
"content": "and ion relaxation time in the low-frequency admittance locus plot of human palmar skin. Medical",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
116,
459,
385,
471
],
"spans": [
{
"bbox": [
116,
459,
385,
471
],
"score": 1.0,
"content": "and Biological Engineering and Computing, 32(2):161–167, 1994.",
"type": "text"
}
],
"index": 27
}
],
"index": 26
},
{
"type": "text",
"bbox": [
107,
478,
505,
512
],
"lines": [
{
"bbox": [
106,
477,
506,
491
],
"spans": [
{
"bbox": [
106,
477,
506,
491
],
"score": 1.0,
"content": "Yao Qin, Nicholas Carlini, Ian J. Goodfellow, Garrison W. Cottrell, and Colin Raffel. Imperceptible,",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
115,
489,
507,
502
],
"spans": [
{
"bbox": [
115,
489,
507,
502
],
"score": 1.0,
"content": "robust, and targeted adversarial examples for automatic speech recognition. CoRR, abs/1903.10346,",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
114,
498,
143,
513
],
"spans": [
{
"bbox": [
114,
498,
143,
513
],
"score": 1.0,
"content": "2019.",
"type": "text"
}
],
"index": 30
}
],
"index": 29
},
{
"type": "text",
"bbox": [
107,
520,
505,
543
],
"lines": [
{
"bbox": [
105,
518,
507,
535
],
"spans": [
{
"bbox": [
105,
518,
507,
535
],
"score": 1.0,
"content": "Mona Sazgar and Michael G. Young. EEG Artifacts, pp. 149–162. Springer International Publishing,",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
115,
530,
171,
544
],
"spans": [
{
"bbox": [
115,
530,
171,
544
],
"score": 1.0,
"content": "Cham, 2019.",
"type": "text"
}
],
"index": 32
}
],
"index": 31.5
},
{
"type": "text",
"bbox": [
107,
550,
506,
596
],
"lines": [
{
"bbox": [
106,
551,
506,
564
],
"spans": [
{
"bbox": [
106,
551,
506,
564
],
"score": 1.0,
"content": "Tibor Schneider, Xiaying Wang, Michael Hersche, Lukas Cavigelli, and Luca Benini. Q-EEGNet: an",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
115,
562,
506,
575
],
"spans": [
{
"bbox": [
115,
562,
506,
575
],
"score": 1.0,
"content": "energy-efficient 8-bit quantized parallel EEGNet implementation for edge motor-imagery brain-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
115,
574,
507,
586
],
"spans": [
{
"bbox": [
115,
574,
507,
586
],
"score": 1.0,
"content": "machine interfaces. In 2020 IEEE International Conference on Smart Computing (SMARTCOMP),",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
115,
585,
196,
596
],
"spans": [
{
"bbox": [
115,
585,
196,
596
],
"score": 1.0,
"content": "pp. 284–289, 2020.",
"type": "text"
}
],
"index": 36
}
],
"index": 34.5
},
{
"type": "text",
"bbox": [
106,
604,
506,
638
],
"lines": [
{
"bbox": [
106,
603,
506,
617
],
"spans": [
{
"bbox": [
106,
603,
506,
617
],
"score": 1.0,
"content": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow,",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
116,
614,
506,
628
],
"spans": [
{
"bbox": [
116,
614,
506,
628
],
"score": 1.0,
"content": "and Rob Fergus. Intriguing properties of neural networks. In Proc. International Conference on",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
115,
626,
276,
638
],
"spans": [
{
"bbox": [
115,
626,
276,
638
],
"score": 1.0,
"content": "Learning Representations, ICLR, 2014.",
"type": "text"
}
],
"index": 39
}
],
"index": 38
},
{
"type": "text",
"bbox": [
106,
645,
505,
679
],
"lines": [
{
"bbox": [
106,
645,
505,
658
],
"spans": [
{
"bbox": [
106,
645,
505,
658
],
"score": 1.0,
"content": "Johannes Vorwerk, Ümit Aydin, Carsten H Wolters, and Christopher R Butson. Influence of head",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
115,
657,
507,
669
],
"spans": [
{
"bbox": [
115,
657,
507,
669
],
"score": 1.0,
"content": "tissue conductivity uncertainties on EEG dipole reconstruction. Frontiers in neuroscience, 13:531,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
115,
667,
143,
680
],
"spans": [
{
"bbox": [
115,
667,
143,
680
],
"score": 1.0,
"content": "2019.",
"type": "text"
}
],
"index": 42
}
],
"index": 41
},
{
"type": "text",
"bbox": [
107,
687,
505,
732
],
"lines": [
{
"bbox": [
106,
687,
505,
699
],
"spans": [
{
"bbox": [
106,
687,
505,
699
],
"score": 1.0,
"content": "Xiaying Wang, Michael Hersche, Batuhan Tömekce, Burak Kaya, Michele Magno, and Luca",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
115,
699,
505,
711
],
"spans": [
{
"bbox": [
115,
699,
505,
711
],
"score": 1.0,
"content": "Benini. An accurate EEGNet-based motor-imagery brain–computer interface for low-power edge",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
114,
709,
505,
723
],
"spans": [
{
"bbox": [
114,
709,
505,
723
],
"score": 1.0,
"content": "computing. In 2020 IEEE International Symposium on Medical Measurements and Applications",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
114,
720,
219,
733
],
"spans": [
{
"bbox": [
114,
720,
219,
733
],
"score": 1.0,
"content": "(MeMeA), pp. 1–6, 2020.",
"type": "text"
}
],
"index": 46
}
],
"index": 44.5
}
],
"page_idx": 11,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
300,
751,
310,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
764
],
"spans": [
{
"bbox": [
299,
750,
312,
764
],
"score": 1.0,
"content": "",
"type": "text",
"height": 14,
"width": 13
}
]
}
]
},
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
105,
82,
504,
105
],
"lines": [
{
"bbox": [
106,
82,
505,
95
],
"spans": [
{
"bbox": [
106,
82,
505,
95
],
"score": 1.0,
"content": "Zihan Liu, Lubin Meng, Xiao Zhang, Weili Fang, and Dongrui Wu. Universal adversarial",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
115,
93,
360,
106
],
"spans": [
{
"bbox": [
115,
93,
360,
106
],
"score": 1.0,
"content": "perturbations for CNN classifiers in EEG-based BCIs, 2021.",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5,
"bbox_fs": [
106,
82,
505,
106
]
},
{
"type": "text",
"bbox": [
108,
113,
504,
147
],
"lines": [
{
"bbox": [
106,
113,
505,
126
],
"spans": [
{
"bbox": [
106,
113,
505,
126
],
"score": 1.0,
"content": "Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu.",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
116,
124,
505,
137
],
"spans": [
{
"bbox": [
116,
124,
505,
137
],
"score": 1.0,
"content": "Towards deep learning models resistant to adversarial attacks. In Proc. International Conference",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
115,
136,
293,
147
],
"spans": [
{
"bbox": [
115,
136,
293,
147
],
"score": 1.0,
"content": "on Learning Representations (ICLR), 2018.",
"type": "text"
}
],
"index": 4
}
],
"index": 3,
"bbox_fs": [
106,
113,
505,
147
]
},
{
"type": "text",
"bbox": [
107,
155,
506,
189
],
"lines": [
{
"bbox": [
106,
155,
505,
167
],
"spans": [
{
"bbox": [
106,
155,
505,
167
],
"score": 1.0,
"content": "Alireza Malekmohammadi, Hoda Mohammadzade, Alireza Chamanzar, Mahdi Shabany, and",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
115,
166,
505,
178
],
"spans": [
{
"bbox": [
115,
166,
505,
178
],
"score": 1.0,
"content": "Benyamin Ghojogh. An efficient hardware implementation for a motor imagery brain computer",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
115,
177,
324,
189
],
"spans": [
{
"bbox": [
115,
177,
324,
189
],
"score": 1.0,
"content": "interface system. Scientia Iranica, 26:72–94, 2019.",
"type": "text"
}
],
"index": 7
}
],
"index": 6,
"bbox_fs": [
106,
155,
505,
189
]
},
{
"type": "text",
"bbox": [
106,
196,
506,
231
],
"lines": [
{
"bbox": [
105,
196,
505,
210
],
"spans": [
{
"bbox": [
105,
196,
505,
210
],
"score": 1.0,
"content": "Iuliana Marin, Myssar Jabbar Hammood Al-Battbootti, and Nicolae Goga. Drone control based on",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
115,
207,
506,
221
],
"spans": [
{
"bbox": [
115,
207,
506,
221
],
"score": 1.0,
"content": "mental commands and facial expressions. In 2020 12th International Conference on Electronics,",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
116,
218,
360,
232
],
"spans": [
{
"bbox": [
116,
218,
360,
232
],
"score": 1.0,
"content": "Computers and Artificial Intelligence (ECAI), pp. 1–4, 2020.",
"type": "text"
}
],
"index": 10
}
],
"index": 9,
"bbox_fs": [
105,
196,
506,
232
]
},
{
"type": "text",
"bbox": [
107,
238,
505,
272
],
"lines": [
{
"bbox": [
105,
237,
506,
252
],
"spans": [
{
"bbox": [
105,
237,
506,
252
],
"score": 1.0,
"content": "Lubin Meng, Jian Huang, Zhigang Zeng, Xue Jiang, Shan Yu, Tzyy-Ping Jung, Chin-Teng Lin,",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
116,
250,
505,
262
],
"spans": [
{
"bbox": [
116,
250,
505,
262
],
"score": 1.0,
"content": "Ricardo Chavarriaga, and Dongrui Wu. EEG-based brain-computer interfaces are vulnerable to",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
117,
261,
214,
273
],
"spans": [
{
"bbox": [
117,
261,
214,
273
],
"score": 1.0,
"content": "backdoor attacks, 2021.",
"type": "text"
}
],
"index": 13
}
],
"index": 12,
"bbox_fs": [
105,
237,
506,
273
]
},
{
"type": "text",
"bbox": [
107,
280,
506,
325
],
"lines": [
{
"bbox": [
105,
279,
507,
293
],
"spans": [
{
"bbox": [
105,
279,
507,
293
],
"score": 1.0,
"content": "Isabelle Merlet, Gwénaël Birot, Ricardo Salvador, Behnam Molaee-Ardekani, Abeye Mekonnen,",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
115,
290,
506,
305
],
"spans": [
{
"bbox": [
115,
290,
506,
305
],
"score": 1.0,
"content": "Aureli Soria-Frish, Giulio Ruffini, Pedro C. Miranda, and Fabrice Wendling. From oscillatory",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
115,
302,
506,
316
],
"spans": [
{
"bbox": [
115,
302,
506,
316
],
"score": 1.0,
"content": "transcranial current stimulation to scalp EEG changes: A biophysical and physiological modeling",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
117,
314,
265,
326
],
"spans": [
{
"bbox": [
117,
314,
265,
326
],
"score": 1.0,
"content": "study. PLOS ONE, 8:1–12, 02 2013.",
"type": "text"
}
],
"index": 17
}
],
"index": 15.5,
"bbox_fs": [
105,
279,
507,
326
]
},
{
"type": "text",
"bbox": [
107,
333,
505,
367
],
"lines": [
{
"bbox": [
106,
333,
505,
345
],
"spans": [
{
"bbox": [
106,
333,
505,
345
],
"score": 1.0,
"content": "Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
115,
344,
505,
357
],
"spans": [
{
"bbox": [
115,
344,
505,
357
],
"score": 1.0,
"content": "adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
115,
355,
248,
369
],
"spans": [
{
"bbox": [
115,
355,
248,
369
],
"score": 1.0,
"content": "Recognition (CVPR), July 2017.",
"type": "text"
}
],
"index": 20
}
],
"index": 19,
"bbox_fs": [
106,
333,
505,
369
]
},
{
"type": "text",
"bbox": [
107,
374,
504,
398
],
"lines": [
{
"bbox": [
106,
375,
505,
387
],
"spans": [
{
"bbox": [
106,
375,
505,
387
],
"score": 1.0,
"content": "G. Pfurtscheller and F.H. Lopes da Silva. Event-related EEG/MEG synchronization and",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
116,
385,
492,
398
],
"spans": [
{
"bbox": [
116,
385,
492,
398
],
"score": 1.0,
"content": "desynchronization: basic principles. Clinical Neurophysiology, 110(11):1842–1857, 11 1999.",
"type": "text"
}
],
"index": 22
}
],
"index": 21.5,
"bbox_fs": [
106,
375,
505,
398
]
},
{
"type": "text",
"bbox": [
106,
406,
505,
428
],
"lines": [
{
"bbox": [
105,
405,
505,
419
],
"spans": [
{
"bbox": [
105,
405,
505,
419
],
"score": 1.0,
"content": "Robert Plutchik and Henry R Hirsch. Skin impedance and phase angle as a function of frequency and",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
115,
417,
295,
428
],
"spans": [
{
"bbox": [
115,
417,
295,
428
],
"score": 1.0,
"content": "current. Science, 141(3584):927–928, 1963.",
"type": "text"
}
],
"index": 24
}
],
"index": 23.5,
"bbox_fs": [
105,
405,
505,
428
]
},
{
"type": "text",
"bbox": [
106,
436,
505,
471
],
"lines": [
{
"bbox": [
105,
435,
505,
450
],
"spans": [
{
"bbox": [
105,
435,
505,
450
],
"score": 1.0,
"content": "ZG Qiao, L Mørkrid, and NK Vøllestad. Characteristics of phase angle, extrapolated conductance",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
117,
449,
505,
460
],
"spans": [
{
"bbox": [
117,
449,
505,
460
],
"score": 1.0,
"content": "and ion relaxation time in the low-frequency admittance locus plot of human palmar skin. Medical",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
116,
459,
385,
471
],
"spans": [
{
"bbox": [
116,
459,
385,
471
],
"score": 1.0,
"content": "and Biological Engineering and Computing, 32(2):161–167, 1994.",
"type": "text"
}
],
"index": 27
}
],
"index": 26,
"bbox_fs": [
105,
435,
505,
471
]
},
{
"type": "text",
"bbox": [
107,
478,
505,
512
],
"lines": [
{
"bbox": [
106,
477,
506,
491
],
"spans": [
{
"bbox": [
106,
477,
506,
491
],
"score": 1.0,
"content": "Yao Qin, Nicholas Carlini, Ian J. Goodfellow, Garrison W. Cottrell, and Colin Raffel. Imperceptible,",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
115,
489,
507,
502
],
"spans": [
{
"bbox": [
115,
489,
507,
502
],
"score": 1.0,
"content": "robust, and targeted adversarial examples for automatic speech recognition. CoRR, abs/1903.10346,",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
114,
498,
143,
513
],
"spans": [
{
"bbox": [
114,
498,
143,
513
],
"score": 1.0,
"content": "2019.",
"type": "text"
}
],
"index": 30
}
],
"index": 29,
"bbox_fs": [
106,
477,
507,
513
]
},
{
"type": "text",
"bbox": [
107,
520,
505,
543
],
"lines": [
{
"bbox": [
105,
518,
507,
535
],
"spans": [
{
"bbox": [
105,
518,
507,
535
],
"score": 1.0,
"content": "Mona Sazgar and Michael G. Young. EEG Artifacts, pp. 149–162. Springer International Publishing,",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
115,
530,
171,
544
],
"spans": [
{
"bbox": [
115,
530,
171,
544
],
"score": 1.0,
"content": "Cham, 2019.",
"type": "text"
}
],
"index": 32
}
],
"index": 31.5,
"bbox_fs": [
105,
518,
507,
544
]
},
{
"type": "text",
"bbox": [
107,
550,
506,
596
],
"lines": [
{
"bbox": [
106,
551,
506,
564
],
"spans": [
{
"bbox": [
106,
551,
506,
564
],
"score": 1.0,
"content": "Tibor Schneider, Xiaying Wang, Michael Hersche, Lukas Cavigelli, and Luca Benini. Q-EEGNet: an",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
115,
562,
506,
575
],
"spans": [
{
"bbox": [
115,
562,
506,
575
],
"score": 1.0,
"content": "energy-efficient 8-bit quantized parallel EEGNet implementation for edge motor-imagery brain-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
115,
574,
507,
586
],
"spans": [
{
"bbox": [
115,
574,
507,
586
],
"score": 1.0,
"content": "machine interfaces. In 2020 IEEE International Conference on Smart Computing (SMARTCOMP),",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
115,
585,
196,
596
],
"spans": [
{
"bbox": [
115,
585,
196,
596
],
"score": 1.0,
"content": "pp. 284–289, 2020.",
"type": "text"
}
],
"index": 36
}
],
"index": 34.5,
"bbox_fs": [
106,
551,
507,
596
]
},
{
"type": "text",
"bbox": [
106,
604,
506,
638
],
"lines": [
{
"bbox": [
106,
603,
506,
617
],
"spans": [
{
"bbox": [
106,
603,
506,
617
],
"score": 1.0,
"content": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow,",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
116,
614,
506,
628
],
"spans": [
{
"bbox": [
116,
614,
506,
628
],
"score": 1.0,
"content": "and Rob Fergus. Intriguing properties of neural networks. In Proc. International Conference on",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
115,
626,
276,
638
],
"spans": [
{
"bbox": [
115,
626,
276,
638
],
"score": 1.0,
"content": "Learning Representations, ICLR, 2014.",
"type": "text"
}
],
"index": 39
}
],
"index": 38,
"bbox_fs": [
106,
603,
506,
638
]
},
{
"type": "text",
"bbox": [
106,
645,
505,
679
],
"lines": [
{
"bbox": [
106,
645,
505,
658
],
"spans": [
{
"bbox": [
106,
645,
505,
658
],
"score": 1.0,
"content": "Johannes Vorwerk, Ümit Aydin, Carsten H Wolters, and Christopher R Butson. Influence of head",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
115,
657,
507,
669
],
"spans": [
{
"bbox": [
115,
657,
507,
669
],
"score": 1.0,
"content": "tissue conductivity uncertainties on EEG dipole reconstruction. Frontiers in neuroscience, 13:531,",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
115,
667,
143,
680
],
"spans": [
{
"bbox": [
115,
667,
143,
680
],
"score": 1.0,
"content": "2019.",
"type": "text"
}
],
"index": 42
}
],
"index": 41,
"bbox_fs": [
106,
645,
507,
680
]
},
{
"type": "text",
"bbox": [
107,
687,
505,
732
],
"lines": [
{
"bbox": [
106,
687,
505,
699
],
"spans": [
{
"bbox": [
106,
687,
505,
699
],
"score": 1.0,
"content": "Xiaying Wang, Michael Hersche, Batuhan Tömekce, Burak Kaya, Michele Magno, and Luca",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
115,
699,
505,
711
],
"spans": [
{
"bbox": [
115,
699,
505,
711
],
"score": 1.0,
"content": "Benini. An accurate EEGNet-based motor-imagery brain–computer interface for low-power edge",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
114,
709,
505,
723
],
"spans": [
{
"bbox": [
114,
709,
505,
723
],
"score": 1.0,
"content": "computing. In 2020 IEEE International Symposium on Medical Measurements and Applications",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
114,
720,
219,
733
],
"spans": [
{
"bbox": [
114,
720,
219,
733
],
"score": 1.0,
"content": "(MeMeA), pp. 1–6, 2020.",
"type": "text"
}
],
"index": 46
}
],
"index": 44.5,
"bbox_fs": [
106,
687,
505,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
108,
82,
504,
116
],
"lines": [
{
"bbox": [
106,
82,
505,
96
],
"spans": [
{
"bbox": [
106,
82,
505,
96
],
"score": 1.0,
"content": "Yang Yu, Yadong Liu, Jun Jiang, Erwei Yin, Zongtan Zhou, and Dewen Hu. An asynchronous control",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
115,
93,
506,
106
],
"spans": [
{
"bbox": [
115,
93,
506,
106
],
"score": 1.0,
"content": "paradigm based on sequential motor imagery and its application in wheelchair navigation. IEEE",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
115,
104,
482,
117
],
"spans": [
{
"bbox": [
115,
104,
482,
117
],
"score": 1.0,
"content": "Transactions on Neural Systems and Rehabilitation Engineering, 26(12):2367–2375, 2018.",
"type": "text"
}
],
"index": 2
}
],
"index": 1
},
{
"type": "text",
"bbox": [
109,
123,
504,
156
],
"lines": [
{
"bbox": [
107,
123,
505,
136
],
"spans": [
{
"bbox": [
107,
123,
505,
136
],
"score": 1.0,
"content": "Wei Emma Zhang, Quan Z. Sheng, Ahoud Abdulrahmn F. Alhazmi, and Chenliang Li. Adversarial",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
115,
133,
506,
147
],
"spans": [
{
"bbox": [
115,
133,
506,
147
],
"score": 1.0,
"content": "attacks on deep-learning models in natural language processing: A survey. ACM Trans. Intell. Syst.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
115,
145,
253,
156
],
"spans": [
{
"bbox": [
115,
145,
253,
156
],
"score": 1.0,
"content": "Technol., 11(3):24:1–24:41, 2020.",
"type": "text"
}
],
"index": 5
}
],
"index": 4
},
{
"type": "text",
"bbox": [
105,
163,
505,
186
],
"lines": [
{
"bbox": [
106,
164,
506,
177
],
"spans": [
{
"bbox": [
106,
164,
506,
177
],
"score": 1.0,
"content": "Xiao Zhang and Dongrui Wu. On the vulnerability of CNN classifiers in EEG-based BCIs. IEEE",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
115,
175,
466,
187
],
"spans": [
{
"bbox": [
115,
175,
466,
187
],
"score": 1.0,
"content": "Transactions on Neural Systems and Rehabilitation Engineering, 27(5):814–825, 2019.",
"type": "text"
}
],
"index": 7
}
],
"index": 6.5
},
{
"type": "text",
"bbox": [
106,
194,
506,
227
],
"lines": [
{
"bbox": [
106,
193,
505,
207
],
"spans": [
{
"bbox": [
106,
193,
505,
207
],
"score": 1.0,
"content": "Jiayu Zhuang, Keke Geng, and Guodong Yin. Ensemble learning based brain–computer interface",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
115,
204,
507,
218
],
"spans": [
{
"bbox": [
115,
204,
507,
218
],
"score": 1.0,
"content": "system for ground vehicle control. IEEE Transactions on Systems, Man, and Cybernetics: Systems,",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
116,
216,
216,
228
],
"spans": [
{
"bbox": [
116,
216,
216,
228
],
"score": 1.0,
"content": "51(9):5392–5404, 2021.",
"type": "text"
}
],
"index": 10
}
],
"index": 9
}
],
"page_idx": 12,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
300,
751,
310,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
764
],
"spans": [
{
"bbox": [
299,
750,
312,
764
],
"score": 1.0,
"content": "",
"type": "text",
"height": 14,
"width": 13
}
]
}
]
},
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
108,
82,
504,
116
],
"lines": [
{
"bbox": [
106,
82,
505,
96
],
"spans": [
{
"bbox": [
106,
82,
505,
96
],
"score": 1.0,
"content": "Yang Yu, Yadong Liu, Jun Jiang, Erwei Yin, Zongtan Zhou, and Dewen Hu. An asynchronous control",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
115,
93,
506,
106
],
"spans": [
{
"bbox": [
115,
93,
506,
106
],
"score": 1.0,
"content": "paradigm based on sequential motor imagery and its application in wheelchair navigation. IEEE",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
115,
104,
482,
117
],
"spans": [
{
"bbox": [
115,
104,
482,
117
],
"score": 1.0,
"content": "Transactions on Neural Systems and Rehabilitation Engineering, 26(12):2367–2375, 2018.",
"type": "text"
}
],
"index": 2
}
],
"index": 1,
"bbox_fs": [
106,
82,
506,
117
]
},
{
"type": "text",
"bbox": [
109,
123,
504,
156
],
"lines": [
{
"bbox": [
107,
123,
505,
136
],
"spans": [
{
"bbox": [
107,
123,
505,
136
],
"score": 1.0,
"content": "Wei Emma Zhang, Quan Z. Sheng, Ahoud Abdulrahmn F. Alhazmi, and Chenliang Li. Adversarial",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
115,
133,
506,
147
],
"spans": [
{
"bbox": [
115,
133,
506,
147
],
"score": 1.0,
"content": "attacks on deep-learning models in natural language processing: A survey. ACM Trans. Intell. Syst.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
115,
145,
253,
156
],
"spans": [
{
"bbox": [
115,
145,
253,
156
],
"score": 1.0,
"content": "Technol., 11(3):24:1–24:41, 2020.",
"type": "text"
}
],
"index": 5
}
],
"index": 4,
"bbox_fs": [
107,
123,
506,
156
]
},
{
"type": "text",
"bbox": [
105,
163,
505,
186
],
"lines": [
{
"bbox": [
106,
164,
506,
177
],
"spans": [
{
"bbox": [
106,
164,
506,
177
],
"score": 1.0,
"content": "Xiao Zhang and Dongrui Wu. On the vulnerability of CNN classifiers in EEG-based BCIs. IEEE",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
115,
175,
466,
187
],
"spans": [
{
"bbox": [
115,
175,
466,
187
],
"score": 1.0,
"content": "Transactions on Neural Systems and Rehabilitation Engineering, 27(5):814–825, 2019.",
"type": "text"
}
],
"index": 7
}
],
"index": 6.5,
"bbox_fs": [
106,
164,
506,
187
]
},
{
"type": "text",
"bbox": [
106,
194,
506,
227
],
"lines": [
{
"bbox": [
106,
193,
505,
207
],
"spans": [
{
"bbox": [
106,
193,
505,
207
],
"score": 1.0,
"content": "Jiayu Zhuang, Keke Geng, and Guodong Yin. Ensemble learning based brain–computer interface",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
115,
204,
507,
218
],
"spans": [
{
"bbox": [
115,
204,
507,
218
],
"score": 1.0,
"content": "system for ground vehicle control. IEEE Transactions on Systems, Man, and Cybernetics: Systems,",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
116,
216,
216,
228
],
"spans": [
{
"bbox": [
116,
216,
216,
228
],
"score": 1.0,
"content": "51(9):5392–5404, 2021.",
"type": "text"
}
],
"index": 10
}
],
"index": 9,
"bbox_fs": [
106,
193,
507,
228
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
110,
89,
499,
148
],
"blocks": [
{
"type": "image_body",
"bbox": [
110,
89,
499,
148
],
"group_id": 1,
"lines": [
{
"bbox": [
110,
89,
499,
148
],
"spans": [
{
"bbox": [
110,
89,
499,
148
],
"score": 0.935,
"type": "image",
"image_path": "673d5d9efd092b4cc8e800ff6a5738b2621fff65b3be06e0ef50d295cf6a25a4.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
110,
89,
499,
108.66666666666667
],
"spans": [],
"index": 0
},
{
"bbox": [
110,
108.66666666666667,
499,
128.33333333333334
],
"spans": [],
"index": 1
},
{
"bbox": [
110,
128.33333333333334,
499,
148.0
],
"spans": [],
"index": 2
}
]
}
],
"index": 1
},
{
"type": "text",
"bbox": [
104,
165,
504,
186
],
"lines": [
{
"bbox": [
106,
165,
505,
177
],
"spans": [
{
"bbox": [
106,
165,
505,
177
],
"score": 1.0,
"content": "Figure 6: Practical adversarial attack scenario in BCIs: a smart device close to the ear emits a perturbation",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
175,
357,
186
],
"spans": [
{
"bbox": [
105,
175,
357,
186
],
"score": 1.0,
"content": "signal which propagates over the head surface to the EEG electrodes.",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5
},
{
"type": "image",
"bbox": [
106,
199,
504,
277
],
"blocks": [
{
"type": "image_body",
"bbox": [
106,
199,
504,
277
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
199,
504,
277
],
"spans": [
{
"bbox": [
106,
199,
504,
277
],
"score": 0.959,
"type": "image",
"image_path": "38b8a7e8b4be2758df986c7eb0277e6018207129bd99c18ee2f96bd9341afb7e.jpg"
}
]
}
],
"index": 6,
"virtual_lines": [
{
"bbox": [
106,
199,
504,
225.0
],
"spans": [],
"index": 5
},
{
"bbox": [
106,
225.0,
504,
251.0
],
"spans": [],
"index": 6
},
{
"bbox": [
106,
251.0,
504,
277.0
],
"spans": [],
"index": 7
}
]
},
{
"type": "image_caption",
"bbox": [
106,
288,
505,
339
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
287,
506,
300
],
"spans": [
{
"bbox": [
105,
287,
324,
300
],
"score": 1.0,
"content": "Figure 7: Magnitude of the spatial propagation for different",
"type": "text"
},
{
"bbox": [
324,
289,
338,
298
],
"score": 0.88,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
338,
287,
506,
300
],
"score": 1.0,
"content": ". The perturbation is emitted from the left side",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
298,
506,
309
],
"spans": [
{
"bbox": [
106,
298,
506,
309
],
"score": 1.0,
"content": "of the head and propagates over the head surface. The leftmost electrode senses the highest magnitude (red),",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
308,
506,
320
],
"spans": [
{
"bbox": [
105,
308,
410,
320
],
"score": 1.0,
"content": "which linearly decreases towards zero (white) with growing propagation distance and",
"type": "text"
},
{
"bbox": [
410,
309,
423,
318
],
"score": 0.88,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
423,
308,
506,
320
],
"score": 1.0,
"content": ". The electrodes which",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
318,
506,
330
],
"spans": [
{
"bbox": [
105,
318,
244,
330
],
"score": 1.0,
"content": "sense the perturbations, i.e., magnitude",
"type": "text"
},
{
"bbox": [
245,
319,
256,
328
],
"score": 0.8,
"content": "{ > } 0",
"type": "inline_equation"
},
{
"bbox": [
256,
318,
506,
330
],
"score": 1.0,
"content": ", are marked with dots. The electrodes T9, C3, C4, and T10 are labeled",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
327,
157,
340
],
"spans": [
{
"bbox": [
106,
327,
157,
340
],
"score": 1.0,
"content": "for reference.",
"type": "text"
}
],
"index": 12
}
],
"index": 10
}
],
"index": 8.0
},
{
"type": "title",
"bbox": [
108,
359,
383,
371
],
"lines": [
{
"bbox": [
105,
358,
383,
374
],
"spans": [
{
"bbox": [
105,
358,
383,
374
],
"score": 1.0,
"content": "A ATTACK AT THE SOURCE OF SIGNAL ACQUISITION",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
106,
383,
504,
406
],
"lines": [
{
"bbox": [
106,
383,
505,
396
],
"spans": [
{
"bbox": [
106,
383,
505,
396
],
"score": 1.0,
"content": "Fig. 6 illustrates the new attack scenario where the perturbation is delivered to the human scalp and",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
394,
406,
407
],
"spans": [
{
"bbox": [
105,
394,
406,
407
],
"score": 1.0,
"content": "propagates to the sensing electrodes at the source of the signal acquisition.",
"type": "text"
}
],
"index": 15
}
],
"index": 14.5
},
{
"type": "title",
"bbox": [
107,
421,
291,
434
],
"lines": [
{
"bbox": [
105,
420,
293,
435
],
"spans": [
{
"bbox": [
105,
420,
293,
435
],
"score": 1.0,
"content": "B SPATIAL PROPAGATION MODELS",
"type": "text"
}
],
"index": 16
}
],
"index": 16
},
{
"type": "text",
"bbox": [
107,
446,
505,
490
],
"lines": [
{
"bbox": [
105,
445,
505,
459
],
"spans": [
{
"bbox": [
105,
445,
478,
459
],
"score": 1.0,
"content": "Fig. 7 illustrates the magnitude of signal propagation using different propagation parameters",
"type": "text"
},
{
"bbox": [
479,
446,
505,
457
],
"score": 0.88,
"content": "\\lambda _ { m } =",
"type": "inline_equation"
}
],
"index": 17
},
{
"bbox": [
107,
457,
506,
470
],
"spans": [
{
"bbox": [
107,
457,
146,
469
],
"score": 0.91,
"content": "\\{ 1 , 5 , 1 5 \\}",
"type": "inline_equation"
},
{
"bbox": [
146,
457,
183,
470
],
"score": 1.0,
"content": ". A large",
"type": "text"
},
{
"bbox": [
183,
457,
198,
468
],
"score": 0.89,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
198,
457,
506,
470
],
"score": 1.0,
"content": "represents cases with large attenuation and limited propagation (e.g., attack",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
468,
505,
481
],
"spans": [
{
"bbox": [
105,
468,
202,
481
],
"score": 1.0,
"content": "over the air) and a small",
"type": "text"
},
{
"bbox": [
202,
469,
217,
479
],
"score": 0.89,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
217,
468,
505,
481
],
"score": 1.0,
"content": "covers cases with lower attenuation where the perturbation can propagate",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
478,
218,
491
],
"spans": [
{
"bbox": [
105,
478,
218,
491
],
"score": 1.0,
"content": "farther (e.g., a smart glass).",
"type": "text"
}
],
"index": 20
}
],
"index": 18.5
},
{
"type": "title",
"bbox": [
107,
506,
355,
519
],
"lines": [
{
"bbox": [
106,
506,
356,
520
],
"spans": [
{
"bbox": [
106,
506,
356,
520
],
"score": 1.0,
"content": "C EXPERIMENTS ON BCI COMPETITION IV-2A",
"type": "text"
}
],
"index": 21
}
],
"index": 21
},
{
"type": "text",
"bbox": [
107,
531,
505,
597
],
"lines": [
{
"bbox": [
106,
531,
505,
543
],
"spans": [
{
"bbox": [
106,
531,
505,
543
],
"score": 1.0,
"content": "Dataset. The IV-2a dataset of the BCI Competition contains recordings from nine different subjects",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
543,
505,
554
],
"spans": [
{
"bbox": [
106,
543,
505,
554
],
"score": 1.0,
"content": "and distinguishes between four classes of imagined movements: left and right hand, both feet, and the",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
553,
506,
565
],
"spans": [
{
"bbox": [
106,
553,
357,
565
],
"score": 1.0,
"content": "tongue. 22 different EEG channels were recorded, sampled at",
"type": "text"
},
{
"bbox": [
357,
553,
387,
564
],
"score": 0.68,
"content": "2 5 0 \\mathrm { H z }",
"type": "inline_equation"
},
{
"bbox": [
387,
553,
506,
565
],
"score": 1.0,
"content": ". The data was pre-processed",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
564,
505,
577
],
"spans": [
{
"bbox": [
106,
564,
260,
577
],
"score": 1.0,
"content": "with a bandpass filter between 0.1 and",
"type": "text"
},
{
"bbox": [
261,
564,
285,
574
],
"score": 0.76,
"content": "4 0 \\mathrm { H z }",
"type": "inline_equation"
},
{
"bbox": [
286,
564,
505,
577
],
"score": 1.0,
"content": ". Each subject completed two recording session on two",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
574,
506,
587
],
"spans": [
{
"bbox": [
105,
574,
506,
587
],
"score": 1.0,
"content": "different days, where the first session is used for training and the second for testing as per the rules of",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
586,
306,
597
],
"spans": [
{
"bbox": [
106,
586,
306,
597
],
"score": 1.0,
"content": "the competition. Each session contains 288 trials.",
"type": "text"
}
],
"index": 27
}
],
"index": 24.5
},
{
"type": "text",
"bbox": [
107,
609,
505,
675
],
"lines": [
{
"bbox": [
106,
609,
506,
622
],
"spans": [
{
"bbox": [
106,
609,
506,
622
],
"score": 1.0,
"content": "Training and validation. We train a separate baseline model per subject using Adam optimizer",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
619,
506,
634
],
"spans": [
{
"bbox": [
105,
619,
126,
634
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
126,
621,
156,
632
],
"score": 0.91,
"content": "\\beta _ { 1 } { = } 0 . 9",
"type": "inline_equation"
},
{
"bbox": [
156,
619,
173,
634
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
174,
621,
213,
632
],
"score": 0.89,
"content": "\\beta _ { 2 } { = } 0 . 9 9 9",
"type": "inline_equation"
},
{
"bbox": [
213,
619,
506,
634
],
"score": 1.0,
"content": ", a batch size of 32, and 500 epochs. The learning rate is 0.001 achieving",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
631,
505,
644
],
"spans": [
{
"bbox": [
105,
631,
199,
644
],
"score": 1.0,
"content": "an average accuracy of",
"type": "text"
},
{
"bbox": [
200,
632,
231,
642
],
"score": 0.88,
"content": "7 1 . 7 9 \\%",
"type": "inline_equation"
},
{
"bbox": [
231,
631,
505,
644
],
"score": 1.0,
"content": ". This dataset does not contain the rest class. We choose to design an",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
642,
506,
655
],
"spans": [
{
"bbox": [
105,
642,
506,
655
],
"score": 1.0,
"content": "attack that aims to fool the classifier to always predict “tongue.” Moreover, we apply a maximum",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
653,
506,
666
],
"spans": [
{
"bbox": [
105,
653,
213,
666
],
"score": 1.0,
"content": "perturbation amplitude of",
"type": "text"
},
{
"bbox": [
213,
653,
285,
665
],
"score": 0.91,
"content": "\\epsilon \\in [ 0 . 0 1 , 1 0 ] \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
286,
653,
506,
666
],
"score": 1.0,
"content": "due to the lower signal amplitude encountered in this",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
663,
140,
677
],
"spans": [
{
"bbox": [
105,
663,
140,
677
],
"score": 1.0,
"content": "dataset.",
"type": "text"
}
],
"index": 33
}
],
"index": 30.5
},
{
"type": "text",
"bbox": [
107,
687,
504,
732
],
"lines": [
{
"bbox": [
105,
687,
505,
700
],
"spans": [
{
"bbox": [
105,
687,
505,
700
],
"score": 1.0,
"content": "Results. Fig. 8 compares the ASR of different attacks without considering the propagation model",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
699,
505,
711
],
"spans": [
{
"bbox": [
105,
699,
338,
711
],
"score": 1.0,
"content": "(Case 1). Generally, a minimal perturbation amplitude of",
"type": "text"
},
{
"bbox": [
339,
699,
361,
709
],
"score": 0.65,
"content": "1 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
362,
699,
379,
711
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
379,
699,
403,
709
],
"score": 0.69,
"content": "2 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
403,
699,
480,
711
],
"score": 1.0,
"content": "suffices to achieve",
"type": "text"
},
{
"bbox": [
480,
699,
505,
709
],
"score": 0.85,
"content": "100 \\%",
"type": "inline_equation"
}
],
"index": 35
},
{
"bbox": [
105,
709,
506,
723
],
"spans": [
{
"bbox": [
105,
709,
506,
723
],
"score": 1.0,
"content": "ASR with PGD and UAP, respectively. The addition of the derivative loss term does not give any",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
720,
506,
734
],
"spans": [
{
"bbox": [
105,
720,
506,
734
],
"score": 1.0,
"content": "performance degradation in terms of the ASR. The average post-attack classification accuracy drops",
"type": "text"
}
],
"index": 37
}
],
"index": 35.5
}
],
"page_idx": 13,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
309,
38
],
"spans": [
{
"bbox": [
106,
25,
309,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
310,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
764
],
"spans": [
{
"bbox": [
299,
750,
312,
764
],
"score": 1.0,
"content": "",
"type": "text",
"height": 14,
"width": 13
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
110,
89,
499,
148
],
"blocks": [
{
"type": "image_body",
"bbox": [
110,
89,
499,
148
],
"group_id": 1,
"lines": [
{
"bbox": [
110,
89,
499,
148
],
"spans": [
{
"bbox": [
110,
89,
499,
148
],
"score": 0.935,
"type": "image",
"image_path": "673d5d9efd092b4cc8e800ff6a5738b2621fff65b3be06e0ef50d295cf6a25a4.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
110,
89,
499,
108.66666666666667
],
"spans": [],
"index": 0
},
{
"bbox": [
110,
108.66666666666667,
499,
128.33333333333334
],
"spans": [],
"index": 1
},
{
"bbox": [
110,
128.33333333333334,
499,
148.0
],
"spans": [],
"index": 2
}
]
}
],
"index": 1
},
{
"type": "text",
"bbox": [
104,
165,
504,
186
],
"lines": [
{
"bbox": [
106,
165,
505,
177
],
"spans": [
{
"bbox": [
106,
165,
505,
177
],
"score": 1.0,
"content": "Figure 6: Practical adversarial attack scenario in BCIs: a smart device close to the ear emits a perturbation",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
175,
357,
186
],
"spans": [
{
"bbox": [
105,
175,
357,
186
],
"score": 1.0,
"content": "signal which propagates over the head surface to the EEG electrodes.",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5,
"bbox_fs": [
105,
165,
505,
186
]
},
{
"type": "image",
"bbox": [
106,
199,
504,
277
],
"blocks": [
{
"type": "image_body",
"bbox": [
106,
199,
504,
277
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
199,
504,
277
],
"spans": [
{
"bbox": [
106,
199,
504,
277
],
"score": 0.959,
"type": "image",
"image_path": "38b8a7e8b4be2758df986c7eb0277e6018207129bd99c18ee2f96bd9341afb7e.jpg"
}
]
}
],
"index": 6,
"virtual_lines": [
{
"bbox": [
106,
199,
504,
225.0
],
"spans": [],
"index": 5
},
{
"bbox": [
106,
225.0,
504,
251.0
],
"spans": [],
"index": 6
},
{
"bbox": [
106,
251.0,
504,
277.0
],
"spans": [],
"index": 7
}
]
},
{
"type": "image_caption",
"bbox": [
106,
288,
505,
339
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
287,
506,
300
],
"spans": [
{
"bbox": [
105,
287,
324,
300
],
"score": 1.0,
"content": "Figure 7: Magnitude of the spatial propagation for different",
"type": "text"
},
{
"bbox": [
324,
289,
338,
298
],
"score": 0.88,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
338,
287,
506,
300
],
"score": 1.0,
"content": ". The perturbation is emitted from the left side",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
106,
298,
506,
309
],
"spans": [
{
"bbox": [
106,
298,
506,
309
],
"score": 1.0,
"content": "of the head and propagates over the head surface. The leftmost electrode senses the highest magnitude (red),",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
308,
506,
320
],
"spans": [
{
"bbox": [
105,
308,
410,
320
],
"score": 1.0,
"content": "which linearly decreases towards zero (white) with growing propagation distance and",
"type": "text"
},
{
"bbox": [
410,
309,
423,
318
],
"score": 0.88,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
423,
308,
506,
320
],
"score": 1.0,
"content": ". The electrodes which",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
318,
506,
330
],
"spans": [
{
"bbox": [
105,
318,
244,
330
],
"score": 1.0,
"content": "sense the perturbations, i.e., magnitude",
"type": "text"
},
{
"bbox": [
245,
319,
256,
328
],
"score": 0.8,
"content": "{ > } 0",
"type": "inline_equation"
},
{
"bbox": [
256,
318,
506,
330
],
"score": 1.0,
"content": ", are marked with dots. The electrodes T9, C3, C4, and T10 are labeled",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
106,
327,
157,
340
],
"spans": [
{
"bbox": [
106,
327,
157,
340
],
"score": 1.0,
"content": "for reference.",
"type": "text"
}
],
"index": 12
}
],
"index": 10
}
],
"index": 8.0
},
{
"type": "title",
"bbox": [
108,
359,
383,
371
],
"lines": [
{
"bbox": [
105,
358,
383,
374
],
"spans": [
{
"bbox": [
105,
358,
383,
374
],
"score": 1.0,
"content": "A ATTACK AT THE SOURCE OF SIGNAL ACQUISITION",
"type": "text"
}
],
"index": 13
}
],
"index": 13
},
{
"type": "text",
"bbox": [
106,
383,
504,
406
],
"lines": [
{
"bbox": [
106,
383,
505,
396
],
"spans": [
{
"bbox": [
106,
383,
505,
396
],
"score": 1.0,
"content": "Fig. 6 illustrates the new attack scenario where the perturbation is delivered to the human scalp and",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
394,
406,
407
],
"spans": [
{
"bbox": [
105,
394,
406,
407
],
"score": 1.0,
"content": "propagates to the sensing electrodes at the source of the signal acquisition.",
"type": "text"
}
],
"index": 15
}
],
"index": 14.5,
"bbox_fs": [
105,
383,
505,
407
]
},
{
"type": "title",
"bbox": [
107,
421,
291,
434
],
"lines": [
{
"bbox": [
105,
420,
293,
435
],
"spans": [
{
"bbox": [
105,
420,
293,
435
],
"score": 1.0,
"content": "B SPATIAL PROPAGATION MODELS",
"type": "text"
}
],
"index": 16
}
],
"index": 16
},
{
"type": "text",
"bbox": [
107,
446,
505,
490
],
"lines": [
{
"bbox": [
105,
445,
505,
459
],
"spans": [
{
"bbox": [
105,
445,
478,
459
],
"score": 1.0,
"content": "Fig. 7 illustrates the magnitude of signal propagation using different propagation parameters",
"type": "text"
},
{
"bbox": [
479,
446,
505,
457
],
"score": 0.88,
"content": "\\lambda _ { m } =",
"type": "inline_equation"
}
],
"index": 17
},
{
"bbox": [
107,
457,
506,
470
],
"spans": [
{
"bbox": [
107,
457,
146,
469
],
"score": 0.91,
"content": "\\{ 1 , 5 , 1 5 \\}",
"type": "inline_equation"
},
{
"bbox": [
146,
457,
183,
470
],
"score": 1.0,
"content": ". A large",
"type": "text"
},
{
"bbox": [
183,
457,
198,
468
],
"score": 0.89,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
198,
457,
506,
470
],
"score": 1.0,
"content": "represents cases with large attenuation and limited propagation (e.g., attack",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
468,
505,
481
],
"spans": [
{
"bbox": [
105,
468,
202,
481
],
"score": 1.0,
"content": "over the air) and a small",
"type": "text"
},
{
"bbox": [
202,
469,
217,
479
],
"score": 0.89,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
217,
468,
505,
481
],
"score": 1.0,
"content": "covers cases with lower attenuation where the perturbation can propagate",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
478,
218,
491
],
"spans": [
{
"bbox": [
105,
478,
218,
491
],
"score": 1.0,
"content": "farther (e.g., a smart glass).",
"type": "text"
}
],
"index": 20
}
],
"index": 18.5,
"bbox_fs": [
105,
445,
506,
491
]
},
{
"type": "title",
"bbox": [
107,
506,
355,
519
],
"lines": [
{
"bbox": [
106,
506,
356,
520
],
"spans": [
{
"bbox": [
106,
506,
356,
520
],
"score": 1.0,
"content": "C EXPERIMENTS ON BCI COMPETITION IV-2A",
"type": "text"
}
],
"index": 21
}
],
"index": 21
},
{
"type": "text",
"bbox": [
107,
531,
505,
597
],
"lines": [
{
"bbox": [
106,
531,
505,
543
],
"spans": [
{
"bbox": [
106,
531,
505,
543
],
"score": 1.0,
"content": "Dataset. The IV-2a dataset of the BCI Competition contains recordings from nine different subjects",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
543,
505,
554
],
"spans": [
{
"bbox": [
106,
543,
505,
554
],
"score": 1.0,
"content": "and distinguishes between four classes of imagined movements: left and right hand, both feet, and the",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
553,
506,
565
],
"spans": [
{
"bbox": [
106,
553,
357,
565
],
"score": 1.0,
"content": "tongue. 22 different EEG channels were recorded, sampled at",
"type": "text"
},
{
"bbox": [
357,
553,
387,
564
],
"score": 0.68,
"content": "2 5 0 \\mathrm { H z }",
"type": "inline_equation"
},
{
"bbox": [
387,
553,
506,
565
],
"score": 1.0,
"content": ". The data was pre-processed",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
564,
505,
577
],
"spans": [
{
"bbox": [
106,
564,
260,
577
],
"score": 1.0,
"content": "with a bandpass filter between 0.1 and",
"type": "text"
},
{
"bbox": [
261,
564,
285,
574
],
"score": 0.76,
"content": "4 0 \\mathrm { H z }",
"type": "inline_equation"
},
{
"bbox": [
286,
564,
505,
577
],
"score": 1.0,
"content": ". Each subject completed two recording session on two",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
574,
506,
587
],
"spans": [
{
"bbox": [
105,
574,
506,
587
],
"score": 1.0,
"content": "different days, where the first session is used for training and the second for testing as per the rules of",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
586,
306,
597
],
"spans": [
{
"bbox": [
106,
586,
306,
597
],
"score": 1.0,
"content": "the competition. Each session contains 288 trials.",
"type": "text"
}
],
"index": 27
}
],
"index": 24.5,
"bbox_fs": [
105,
531,
506,
597
]
},
{
"type": "text",
"bbox": [
107,
609,
505,
675
],
"lines": [
{
"bbox": [
106,
609,
506,
622
],
"spans": [
{
"bbox": [
106,
609,
506,
622
],
"score": 1.0,
"content": "Training and validation. We train a separate baseline model per subject using Adam optimizer",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
619,
506,
634
],
"spans": [
{
"bbox": [
105,
619,
126,
634
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
126,
621,
156,
632
],
"score": 0.91,
"content": "\\beta _ { 1 } { = } 0 . 9",
"type": "inline_equation"
},
{
"bbox": [
156,
619,
173,
634
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
174,
621,
213,
632
],
"score": 0.89,
"content": "\\beta _ { 2 } { = } 0 . 9 9 9",
"type": "inline_equation"
},
{
"bbox": [
213,
619,
506,
634
],
"score": 1.0,
"content": ", a batch size of 32, and 500 epochs. The learning rate is 0.001 achieving",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
631,
505,
644
],
"spans": [
{
"bbox": [
105,
631,
199,
644
],
"score": 1.0,
"content": "an average accuracy of",
"type": "text"
},
{
"bbox": [
200,
632,
231,
642
],
"score": 0.88,
"content": "7 1 . 7 9 \\%",
"type": "inline_equation"
},
{
"bbox": [
231,
631,
505,
644
],
"score": 1.0,
"content": ". This dataset does not contain the rest class. We choose to design an",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
642,
506,
655
],
"spans": [
{
"bbox": [
105,
642,
506,
655
],
"score": 1.0,
"content": "attack that aims to fool the classifier to always predict “tongue.” Moreover, we apply a maximum",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
653,
506,
666
],
"spans": [
{
"bbox": [
105,
653,
213,
666
],
"score": 1.0,
"content": "perturbation amplitude of",
"type": "text"
},
{
"bbox": [
213,
653,
285,
665
],
"score": 0.91,
"content": "\\epsilon \\in [ 0 . 0 1 , 1 0 ] \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
286,
653,
506,
666
],
"score": 1.0,
"content": "due to the lower signal amplitude encountered in this",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
663,
140,
677
],
"spans": [
{
"bbox": [
105,
663,
140,
677
],
"score": 1.0,
"content": "dataset.",
"type": "text"
}
],
"index": 33
}
],
"index": 30.5,
"bbox_fs": [
105,
609,
506,
677
]
},
{
"type": "text",
"bbox": [
107,
687,
504,
732
],
"lines": [
{
"bbox": [
105,
687,
505,
700
],
"spans": [
{
"bbox": [
105,
687,
505,
700
],
"score": 1.0,
"content": "Results. Fig. 8 compares the ASR of different attacks without considering the propagation model",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
699,
505,
711
],
"spans": [
{
"bbox": [
105,
699,
338,
711
],
"score": 1.0,
"content": "(Case 1). Generally, a minimal perturbation amplitude of",
"type": "text"
},
{
"bbox": [
339,
699,
361,
709
],
"score": 0.65,
"content": "1 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
362,
699,
379,
711
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
379,
699,
403,
709
],
"score": 0.69,
"content": "2 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
403,
699,
480,
711
],
"score": 1.0,
"content": "suffices to achieve",
"type": "text"
},
{
"bbox": [
480,
699,
505,
709
],
"score": 0.85,
"content": "100 \\%",
"type": "inline_equation"
}
],
"index": 35
},
{
"bbox": [
105,
709,
506,
723
],
"spans": [
{
"bbox": [
105,
709,
506,
723
],
"score": 1.0,
"content": "ASR with PGD and UAP, respectively. The addition of the derivative loss term does not give any",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
720,
506,
734
],
"spans": [
{
"bbox": [
105,
720,
506,
734
],
"score": 1.0,
"content": "performance degradation in terms of the ASR. The average post-attack classification accuracy drops",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
332,
504,
347
],
"spans": [
{
"bbox": [
105,
332,
129,
347
],
"score": 1.0,
"content": "from",
"type": "text",
"cross_page": true
},
{
"bbox": [
129,
334,
162,
344
],
"score": 0.86,
"content": "7 1 . 7 9 \\%",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
162,
332,
174,
347
],
"score": 1.0,
"content": "to",
"type": "text",
"cross_page": true
},
{
"bbox": [
174,
334,
195,
344
],
"score": 0.86,
"content": "50 \\%",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
195,
332,
326,
347
],
"score": 1.0,
"content": "for a perturbation amplitude of",
"type": "text",
"cross_page": true
},
{
"bbox": [
327,
334,
363,
344
],
"score": 0.8,
"content": "0 . 1 5 \\mathrm { m V }",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
363,
332,
382,
347
],
"score": 1.0,
"content": "and",
"type": "text",
"cross_page": true
},
{
"bbox": [
382,
334,
410,
344
],
"score": 0.86,
"content": "2 4 . 7 \\%",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
410,
332,
426,
347
],
"score": 1.0,
"content": "for",
"type": "text",
"cross_page": true
},
{
"bbox": [
426,
334,
457,
344
],
"score": 0.82,
"content": "0 . 6 \\mathrm { m V }",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
458,
332,
504,
347
],
"score": 1.0,
"content": "and higher",
"type": "text",
"cross_page": true
}
],
"index": 5
},
{
"bbox": [
106,
344,
295,
357
],
"spans": [
{
"bbox": [
106,
344,
295,
357
],
"score": 1.0,
"content": "amplitudes, when PGD with derivative is used.",
"type": "text",
"cross_page": true
}
],
"index": 6
}
],
"index": 35.5,
"bbox_fs": [
105,
687,
506,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
111,
80,
495,
262
],
"blocks": [
{
"type": "image_body",
"bbox": [
111,
80,
495,
262
],
"group_id": 0,
"lines": [
{
"bbox": [
111,
80,
495,
262
],
"spans": [
{
"bbox": [
111,
80,
495,
262
],
"score": 0.784,
"type": "image",
"image_path": "2754d89a4671045e54c9d9a042e28c5bfceebb00058d60cea27bd8868f79baae.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
111,
80,
495,
140.66666666666666
],
"spans": [],
"index": 0
},
{
"bbox": [
111,
140.66666666666666,
495,
201.33333333333331
],
"spans": [],
"index": 1
},
{
"bbox": [
111,
201.33333333333331,
495,
262.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
105,
273,
504,
295
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
273,
506,
285
],
"spans": [
{
"bbox": [
105,
273,
506,
285
],
"score": 1.0,
"content": "Figure 8: ASR on BCI Competition IV-2a with random noise, FGSM, PGD, and UAP with and without",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
283,
181,
295
],
"spans": [
{
"bbox": [
105,
283,
181,
295
],
"score": 1.0,
"content": "derivative loss term.",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5
}
],
"index": 2.25
},
{
"type": "text",
"bbox": [
107,
333,
502,
356
],
"lines": [
{
"bbox": [
105,
332,
504,
347
],
"spans": [
{
"bbox": [
105,
332,
129,
347
],
"score": 1.0,
"content": "from",
"type": "text"
},
{
"bbox": [
129,
334,
162,
344
],
"score": 0.86,
"content": "7 1 . 7 9 \\%",
"type": "inline_equation"
},
{
"bbox": [
162,
332,
174,
347
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
174,
334,
195,
344
],
"score": 0.86,
"content": "50 \\%",
"type": "inline_equation"
},
{
"bbox": [
195,
332,
326,
347
],
"score": 1.0,
"content": "for a perturbation amplitude of",
"type": "text"
},
{
"bbox": [
327,
334,
363,
344
],
"score": 0.8,
"content": "0 . 1 5 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
363,
332,
382,
347
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
382,
334,
410,
344
],
"score": 0.86,
"content": "2 4 . 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
410,
332,
426,
347
],
"score": 1.0,
"content": "for",
"type": "text"
},
{
"bbox": [
426,
334,
457,
344
],
"score": 0.82,
"content": "0 . 6 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
458,
332,
504,
347
],
"score": 1.0,
"content": "and higher",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
344,
295,
357
],
"spans": [
{
"bbox": [
106,
344,
295,
357
],
"score": 1.0,
"content": "amplitudes, when PGD with derivative is used.",
"type": "text"
}
],
"index": 6
}
],
"index": 5.5
},
{
"type": "text",
"bbox": [
107,
361,
505,
405
],
"lines": [
{
"bbox": [
105,
361,
505,
374
],
"spans": [
{
"bbox": [
105,
361,
345,
374
],
"score": 1.0,
"content": "Fig. 9 shows the ASR for different propagation parameters",
"type": "text"
},
{
"bbox": [
345,
362,
360,
372
],
"score": 0.88,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
361,
361,
378,
374
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
378,
362,
390,
372
],
"score": 0.84,
"content": "\\lambda _ { d , }",
"type": "inline_equation"
},
{
"bbox": [
391,
361,
505,
374
],
"score": 1.0,
"content": ") and maximum perturbation",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
372,
505,
384
],
"spans": [
{
"bbox": [
105,
372,
151,
384
],
"score": 1.0,
"content": "amplitudes",
"type": "text"
},
{
"bbox": [
151,
374,
157,
382
],
"score": 0.53,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
158,
372,
505,
384
],
"score": 1.0,
"content": ". When considering the head model during the design of the attack (Case 2, w/HM), both",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
383,
505,
396
],
"spans": [
{
"bbox": [
105,
383,
505,
396
],
"score": 1.0,
"content": "PGD and UAP reach significantly higher ASR compared to attacks designed without the consideration",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
394,
255,
406
],
"spans": [
{
"bbox": [
105,
394,
255,
406
],
"score": 1.0,
"content": "of the head model (Case 3, w/oHM).",
"type": "text"
}
],
"index": 10
}
],
"index": 8.5
},
{
"type": "title",
"bbox": [
108,
442,
266,
454
],
"lines": [
{
"bbox": [
105,
441,
268,
456
],
"spans": [
{
"bbox": [
105,
441,
268,
456
],
"score": 1.0,
"content": "D PLAUSIBILITY OF ATTACKS",
"type": "text"
}
],
"index": 11
}
],
"index": 11
},
{
"type": "text",
"bbox": [
106,
478,
505,
567
],
"lines": [
{
"bbox": [
105,
478,
505,
492
],
"spans": [
{
"bbox": [
105,
478,
505,
492
],
"score": 1.0,
"content": "This section provides power spectral density plots of original signals and attacked signals with and",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
106,
489,
505,
501
],
"spans": [
{
"bbox": [
106,
489,
505,
501
],
"score": 1.0,
"content": "without the derivative loss term, shown in Figure 10. The power spectral density is determined",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
501,
505,
513
],
"spans": [
{
"bbox": [
106,
501,
505,
513
],
"score": 1.0,
"content": "by computing the magnitude squared Fast Fourier Transform of the signals that were illustrated in",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
511,
505,
524
],
"spans": [
{
"bbox": [
105,
511,
505,
524
],
"score": 1.0,
"content": "Figure 2. The attack designed with the derivative loss term has a similar distribution as the original",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
523,
505,
535
],
"spans": [
{
"bbox": [
105,
523,
505,
535
],
"score": 1.0,
"content": "signal, where as the attack without derivative shows large contributions in the low frequency domain",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
107,
533,
505,
546
],
"spans": [
{
"bbox": [
107,
533,
139,
545
],
"score": 0.85,
"content": "( < 5 \\mathrm { H z } )",
"type": "inline_equation"
},
{
"bbox": [
139,
533,
505,
546
],
"score": 1.0,
"content": ", which were not present in the original signal. These low-frequency components stem from",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
545,
506,
557
],
"spans": [
{
"bbox": [
106,
545,
506,
557
],
"score": 1.0,
"content": "the square-wave shaped attack and can be used as a way to detect the attack; hence, this attack cannot",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
556,
224,
568
],
"spans": [
{
"bbox": [
106,
556,
224,
568
],
"score": 1.0,
"content": "be considered imperceptible.",
"type": "text"
}
],
"index": 19
}
],
"index": 15.5
},
{
"type": "text",
"bbox": [
106,
572,
505,
671
],
"lines": [
{
"bbox": [
106,
572,
505,
584
],
"spans": [
{
"bbox": [
106,
572,
505,
584
],
"score": 1.0,
"content": "Moreover, Figure 13 shows the attacks with and without derivative loss term with increasing maximum",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
582,
505,
595
],
"spans": [
{
"bbox": [
105,
582,
309,
595
],
"score": 1.0,
"content": "amplitude \u000f. We can see that for low amplitudes (",
"type": "text"
},
{
"bbox": [
309,
583,
333,
594
],
"score": 0.56,
"content": "\\mathrm { 1 m V }",
"type": "inline_equation"
},
{
"bbox": [
333,
582,
351,
595
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
351,
583,
375,
594
],
"score": 0.47,
"content": "5 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
375,
582,
505,
595
],
"score": 1.0,
"content": ") the generated attacks with and",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
593,
505,
607
],
"spans": [
{
"bbox": [
105,
593,
280,
607
],
"score": 1.0,
"content": "without derivative still look like EEGs. At",
"type": "text"
},
{
"bbox": [
280,
594,
307,
605
],
"score": 0.7,
"content": "1 0 \\mathrm { m V } ,",
"type": "inline_equation"
},
{
"bbox": [
308,
593,
505,
607
],
"score": 1.0,
"content": "the attack generated without derivative presents",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
604,
506,
618
],
"spans": [
{
"bbox": [
105,
604,
457,
618
],
"score": 1.0,
"content": "minor square-wave artifacts, which could be still imperceptible to a non-expert. With",
"type": "text"
},
{
"bbox": [
458,
605,
487,
615
],
"score": 0.73,
"content": "2 5 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
487,
604,
506,
618
],
"score": 1.0,
"content": "and",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
615,
506,
629
],
"spans": [
{
"bbox": [
106,
616,
134,
627
],
"score": 0.7,
"content": "5 0 \\mathrm { m V } ,",
"type": "inline_equation"
},
{
"bbox": [
134,
615,
506,
629
],
"score": 1.0,
"content": ", the ones generated without derivative have strong and perceptible square-wave displacements,",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
627,
505,
639
],
"spans": [
{
"bbox": [
106,
627,
505,
639
],
"score": 1.0,
"content": "while the ones generated with our proposed method can still be mistaken as real EEG signals. While",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
638,
504,
650
],
"spans": [
{
"bbox": [
106,
638,
382,
650
],
"score": 1.0,
"content": "with the instance-based attacks, it is not necessary to have more than",
"type": "text"
},
{
"bbox": [
382,
638,
410,
648
],
"score": 0.75,
"content": "1 0 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
410,
638,
504,
650
],
"score": 1.0,
"content": "to get a very high ASR",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
104,
647,
506,
663
],
"spans": [
{
"bbox": [
104,
647,
506,
663
],
"score": 1.0,
"content": "(see Figure 1), with the universal attacks and physical constraints, the ASR increases with increasing",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
660,
260,
673
],
"spans": [
{
"bbox": [
105,
660,
260,
673
],
"score": 1.0,
"content": "perturbation amplitude (see Figure 4).",
"type": "text"
}
],
"index": 28
}
],
"index": 24
},
{
"type": "text",
"bbox": [
107,
677,
505,
732
],
"lines": [
{
"bbox": [
106,
677,
505,
689
],
"spans": [
{
"bbox": [
106,
677,
505,
689
],
"score": 1.0,
"content": "The same observations can be drawn from the plausibility metrics, which have been proposed for the",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
688,
505,
700
],
"spans": [
{
"bbox": [
106,
688,
505,
700
],
"score": 1.0,
"content": "first time in this paper to assess quantitatively the EEG attacks. For example, looking at the cosine",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
697,
507,
713
],
"spans": [
{
"bbox": [
105,
697,
147,
713
],
"score": 1.0,
"content": "similarity",
"type": "text"
},
{
"bbox": [
147,
699,
160,
710
],
"score": 0.8,
"content": "( \\gamma )",
"type": "inline_equation"
},
{
"bbox": [
161,
697,
335,
713
],
"score": 1.0,
"content": "in Table 1, without the derivative loss term,",
"type": "text"
},
{
"bbox": [
335,
700,
343,
710
],
"score": 0.8,
"content": "\\gamma",
"type": "inline_equation"
},
{
"bbox": [
343,
697,
378,
713
],
"score": 1.0,
"content": "drops to",
"type": "text"
},
{
"bbox": [
378,
699,
410,
709
],
"score": 0.88,
"content": "9 7 . 9 9 \\%",
"type": "inline_equation"
},
{
"bbox": [
411,
697,
431,
713
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
432,
699,
466,
709
],
"score": 0.89,
"content": "\\epsilon = 5 \\mathrm { m V } ,",
"type": "inline_equation"
},
{
"bbox": [
466,
697,
507,
713
],
"score": 1.0,
"content": ", whereas,",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
708,
505,
723
],
"spans": [
{
"bbox": [
105,
708,
184,
723
],
"score": 1.0,
"content": "with the derivative,",
"type": "text"
},
{
"bbox": [
185,
711,
192,
721
],
"score": 0.8,
"content": "\\gamma",
"type": "inline_equation"
},
{
"bbox": [
192,
708,
324,
723
],
"score": 1.0,
"content": "drops to about the same value of",
"type": "text"
},
{
"bbox": [
325,
710,
357,
720
],
"score": 0.88,
"content": "9 7 . 4 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
357,
708,
378,
723
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
378,
710,
420,
720
],
"score": 0.89,
"content": "\\epsilon = 1 0 \\mathrm { m V } ,",
"type": "inline_equation"
},
{
"bbox": [
421,
708,
505,
723
],
"score": 1.0,
"content": "yielding an increase",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
720,
398,
733
],
"spans": [
{
"bbox": [
105,
720,
160,
733
],
"score": 1.0,
"content": "in ASR from",
"type": "text"
},
{
"bbox": [
160,
721,
180,
731
],
"score": 0.82,
"content": "85 \\%",
"type": "inline_equation"
},
{
"bbox": [
180,
720,
184,
733
],
"score": 0.0,
"content": "",
"type": "text"
},
{
"bbox": [
184,
721,
209,
731
],
"score": 0.41,
"content": "\\mathrm { 5 m V ) }",
"type": "inline_equation"
},
{
"bbox": [
210,
720,
222,
733
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
222,
721,
242,
731
],
"score": 0.83,
"content": "9 9 \\%",
"type": "inline_equation"
},
{
"bbox": [
242,
720,
246,
733
],
"score": 0.0,
"content": "",
"type": "text"
},
{
"bbox": [
246,
721,
276,
731
],
"score": 0.62,
"content": "\\mathrm { 1 0 m V ) }",
"type": "inline_equation"
},
{
"bbox": [
276,
720,
398,
733
],
"score": 1.0,
"content": "shown in Figure 1 with PGD.",
"type": "text"
}
],
"index": 33
}
],
"index": 31
}
],
"page_idx": 14,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
26,
309,
38
],
"spans": [
{
"bbox": [
106,
26,
309,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
310,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
764
],
"spans": [
{
"bbox": [
299,
750,
312,
764
],
"score": 1.0,
"content": "",
"type": "text",
"height": 14,
"width": 13
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
111,
80,
495,
262
],
"blocks": [
{
"type": "image_body",
"bbox": [
111,
80,
495,
262
],
"group_id": 0,
"lines": [
{
"bbox": [
111,
80,
495,
262
],
"spans": [
{
"bbox": [
111,
80,
495,
262
],
"score": 0.784,
"type": "image",
"image_path": "2754d89a4671045e54c9d9a042e28c5bfceebb00058d60cea27bd8868f79baae.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
111,
80,
495,
140.66666666666666
],
"spans": [],
"index": 0
},
{
"bbox": [
111,
140.66666666666666,
495,
201.33333333333331
],
"spans": [],
"index": 1
},
{
"bbox": [
111,
201.33333333333331,
495,
262.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
105,
273,
504,
295
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
273,
506,
285
],
"spans": [
{
"bbox": [
105,
273,
506,
285
],
"score": 1.0,
"content": "Figure 8: ASR on BCI Competition IV-2a with random noise, FGSM, PGD, and UAP with and without",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
283,
181,
295
],
"spans": [
{
"bbox": [
105,
283,
181,
295
],
"score": 1.0,
"content": "derivative loss term.",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5
}
],
"index": 2.25
},
{
"type": "text",
"bbox": [
107,
333,
502,
356
],
"lines": [],
"index": 5.5,
"bbox_fs": [
105,
332,
504,
357
],
"lines_deleted": true
},
{
"type": "text",
"bbox": [
107,
361,
505,
405
],
"lines": [
{
"bbox": [
105,
361,
505,
374
],
"spans": [
{
"bbox": [
105,
361,
345,
374
],
"score": 1.0,
"content": "Fig. 9 shows the ASR for different propagation parameters",
"type": "text"
},
{
"bbox": [
345,
362,
360,
372
],
"score": 0.88,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
361,
361,
378,
374
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
378,
362,
390,
372
],
"score": 0.84,
"content": "\\lambda _ { d , }",
"type": "inline_equation"
},
{
"bbox": [
391,
361,
505,
374
],
"score": 1.0,
"content": ") and maximum perturbation",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
372,
505,
384
],
"spans": [
{
"bbox": [
105,
372,
151,
384
],
"score": 1.0,
"content": "amplitudes",
"type": "text"
},
{
"bbox": [
151,
374,
157,
382
],
"score": 0.53,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
158,
372,
505,
384
],
"score": 1.0,
"content": ". When considering the head model during the design of the attack (Case 2, w/HM), both",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
383,
505,
396
],
"spans": [
{
"bbox": [
105,
383,
505,
396
],
"score": 1.0,
"content": "PGD and UAP reach significantly higher ASR compared to attacks designed without the consideration",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
394,
255,
406
],
"spans": [
{
"bbox": [
105,
394,
255,
406
],
"score": 1.0,
"content": "of the head model (Case 3, w/oHM).",
"type": "text"
}
],
"index": 10
}
],
"index": 8.5,
"bbox_fs": [
105,
361,
505,
406
]
},
{
"type": "title",
"bbox": [
108,
442,
266,
454
],
"lines": [
{
"bbox": [
105,
441,
268,
456
],
"spans": [
{
"bbox": [
105,
441,
268,
456
],
"score": 1.0,
"content": "D PLAUSIBILITY OF ATTACKS",
"type": "text"
}
],
"index": 11
}
],
"index": 11
},
{
"type": "text",
"bbox": [
106,
478,
505,
567
],
"lines": [
{
"bbox": [
105,
478,
505,
492
],
"spans": [
{
"bbox": [
105,
478,
505,
492
],
"score": 1.0,
"content": "This section provides power spectral density plots of original signals and attacked signals with and",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
106,
489,
505,
501
],
"spans": [
{
"bbox": [
106,
489,
505,
501
],
"score": 1.0,
"content": "without the derivative loss term, shown in Figure 10. The power spectral density is determined",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
501,
505,
513
],
"spans": [
{
"bbox": [
106,
501,
505,
513
],
"score": 1.0,
"content": "by computing the magnitude squared Fast Fourier Transform of the signals that were illustrated in",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
511,
505,
524
],
"spans": [
{
"bbox": [
105,
511,
505,
524
],
"score": 1.0,
"content": "Figure 2. The attack designed with the derivative loss term has a similar distribution as the original",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
523,
505,
535
],
"spans": [
{
"bbox": [
105,
523,
505,
535
],
"score": 1.0,
"content": "signal, where as the attack without derivative shows large contributions in the low frequency domain",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
107,
533,
505,
546
],
"spans": [
{
"bbox": [
107,
533,
139,
545
],
"score": 0.85,
"content": "( < 5 \\mathrm { H z } )",
"type": "inline_equation"
},
{
"bbox": [
139,
533,
505,
546
],
"score": 1.0,
"content": ", which were not present in the original signal. These low-frequency components stem from",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
545,
506,
557
],
"spans": [
{
"bbox": [
106,
545,
506,
557
],
"score": 1.0,
"content": "the square-wave shaped attack and can be used as a way to detect the attack; hence, this attack cannot",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
556,
224,
568
],
"spans": [
{
"bbox": [
106,
556,
224,
568
],
"score": 1.0,
"content": "be considered imperceptible.",
"type": "text"
}
],
"index": 19
}
],
"index": 15.5,
"bbox_fs": [
105,
478,
506,
568
]
},
{
"type": "text",
"bbox": [
106,
572,
505,
671
],
"lines": [
{
"bbox": [
106,
572,
505,
584
],
"spans": [
{
"bbox": [
106,
572,
505,
584
],
"score": 1.0,
"content": "Moreover, Figure 13 shows the attacks with and without derivative loss term with increasing maximum",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
582,
505,
595
],
"spans": [
{
"bbox": [
105,
582,
309,
595
],
"score": 1.0,
"content": "amplitude \u000f. We can see that for low amplitudes (",
"type": "text"
},
{
"bbox": [
309,
583,
333,
594
],
"score": 0.56,
"content": "\\mathrm { 1 m V }",
"type": "inline_equation"
},
{
"bbox": [
333,
582,
351,
595
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
351,
583,
375,
594
],
"score": 0.47,
"content": "5 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
375,
582,
505,
595
],
"score": 1.0,
"content": ") the generated attacks with and",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
593,
505,
607
],
"spans": [
{
"bbox": [
105,
593,
280,
607
],
"score": 1.0,
"content": "without derivative still look like EEGs. At",
"type": "text"
},
{
"bbox": [
280,
594,
307,
605
],
"score": 0.7,
"content": "1 0 \\mathrm { m V } ,",
"type": "inline_equation"
},
{
"bbox": [
308,
593,
505,
607
],
"score": 1.0,
"content": "the attack generated without derivative presents",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
604,
506,
618
],
"spans": [
{
"bbox": [
105,
604,
457,
618
],
"score": 1.0,
"content": "minor square-wave artifacts, which could be still imperceptible to a non-expert. With",
"type": "text"
},
{
"bbox": [
458,
605,
487,
615
],
"score": 0.73,
"content": "2 5 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
487,
604,
506,
618
],
"score": 1.0,
"content": "and",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
615,
506,
629
],
"spans": [
{
"bbox": [
106,
616,
134,
627
],
"score": 0.7,
"content": "5 0 \\mathrm { m V } ,",
"type": "inline_equation"
},
{
"bbox": [
134,
615,
506,
629
],
"score": 1.0,
"content": ", the ones generated without derivative have strong and perceptible square-wave displacements,",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
627,
505,
639
],
"spans": [
{
"bbox": [
106,
627,
505,
639
],
"score": 1.0,
"content": "while the ones generated with our proposed method can still be mistaken as real EEG signals. While",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
638,
504,
650
],
"spans": [
{
"bbox": [
106,
638,
382,
650
],
"score": 1.0,
"content": "with the instance-based attacks, it is not necessary to have more than",
"type": "text"
},
{
"bbox": [
382,
638,
410,
648
],
"score": 0.75,
"content": "1 0 \\mathrm { m V }",
"type": "inline_equation"
},
{
"bbox": [
410,
638,
504,
650
],
"score": 1.0,
"content": "to get a very high ASR",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
104,
647,
506,
663
],
"spans": [
{
"bbox": [
104,
647,
506,
663
],
"score": 1.0,
"content": "(see Figure 1), with the universal attacks and physical constraints, the ASR increases with increasing",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
660,
260,
673
],
"spans": [
{
"bbox": [
105,
660,
260,
673
],
"score": 1.0,
"content": "perturbation amplitude (see Figure 4).",
"type": "text"
}
],
"index": 28
}
],
"index": 24,
"bbox_fs": [
104,
572,
506,
673
]
},
{
"type": "text",
"bbox": [
107,
677,
505,
732
],
"lines": [
{
"bbox": [
106,
677,
505,
689
],
"spans": [
{
"bbox": [
106,
677,
505,
689
],
"score": 1.0,
"content": "The same observations can be drawn from the plausibility metrics, which have been proposed for the",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
688,
505,
700
],
"spans": [
{
"bbox": [
106,
688,
505,
700
],
"score": 1.0,
"content": "first time in this paper to assess quantitatively the EEG attacks. For example, looking at the cosine",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
697,
507,
713
],
"spans": [
{
"bbox": [
105,
697,
147,
713
],
"score": 1.0,
"content": "similarity",
"type": "text"
},
{
"bbox": [
147,
699,
160,
710
],
"score": 0.8,
"content": "( \\gamma )",
"type": "inline_equation"
},
{
"bbox": [
161,
697,
335,
713
],
"score": 1.0,
"content": "in Table 1, without the derivative loss term,",
"type": "text"
},
{
"bbox": [
335,
700,
343,
710
],
"score": 0.8,
"content": "\\gamma",
"type": "inline_equation"
},
{
"bbox": [
343,
697,
378,
713
],
"score": 1.0,
"content": "drops to",
"type": "text"
},
{
"bbox": [
378,
699,
410,
709
],
"score": 0.88,
"content": "9 7 . 9 9 \\%",
"type": "inline_equation"
},
{
"bbox": [
411,
697,
431,
713
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
432,
699,
466,
709
],
"score": 0.89,
"content": "\\epsilon = 5 \\mathrm { m V } ,",
"type": "inline_equation"
},
{
"bbox": [
466,
697,
507,
713
],
"score": 1.0,
"content": ", whereas,",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
708,
505,
723
],
"spans": [
{
"bbox": [
105,
708,
184,
723
],
"score": 1.0,
"content": "with the derivative,",
"type": "text"
},
{
"bbox": [
185,
711,
192,
721
],
"score": 0.8,
"content": "\\gamma",
"type": "inline_equation"
},
{
"bbox": [
192,
708,
324,
723
],
"score": 1.0,
"content": "drops to about the same value of",
"type": "text"
},
{
"bbox": [
325,
710,
357,
720
],
"score": 0.88,
"content": "9 7 . 4 7 \\%",
"type": "inline_equation"
},
{
"bbox": [
357,
708,
378,
723
],
"score": 1.0,
"content": "with",
"type": "text"
},
{
"bbox": [
378,
710,
420,
720
],
"score": 0.89,
"content": "\\epsilon = 1 0 \\mathrm { m V } ,",
"type": "inline_equation"
},
{
"bbox": [
421,
708,
505,
723
],
"score": 1.0,
"content": "yielding an increase",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
720,
398,
733
],
"spans": [
{
"bbox": [
105,
720,
160,
733
],
"score": 1.0,
"content": "in ASR from",
"type": "text"
},
{
"bbox": [
160,
721,
180,
731
],
"score": 0.82,
"content": "85 \\%",
"type": "inline_equation"
},
{
"bbox": [
180,
720,
184,
733
],
"score": 0.0,
"content": "",
"type": "text"
},
{
"bbox": [
184,
721,
209,
731
],
"score": 0.41,
"content": "\\mathrm { 5 m V ) }",
"type": "inline_equation"
},
{
"bbox": [
210,
720,
222,
733
],
"score": 1.0,
"content": "to",
"type": "text"
},
{
"bbox": [
222,
721,
242,
731
],
"score": 0.83,
"content": "9 9 \\%",
"type": "inline_equation"
},
{
"bbox": [
242,
720,
246,
733
],
"score": 0.0,
"content": "",
"type": "text"
},
{
"bbox": [
246,
721,
276,
731
],
"score": 0.62,
"content": "\\mathrm { 1 0 m V ) }",
"type": "inline_equation"
},
{
"bbox": [
276,
720,
398,
733
],
"score": 1.0,
"content": "shown in Figure 1 with PGD.",
"type": "text"
}
],
"index": 33
}
],
"index": 31,
"bbox_fs": [
105,
677,
507,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
169,
82,
442,
329
],
"blocks": [
{
"type": "image_body",
"bbox": [
169,
82,
442,
329
],
"group_id": 0,
"lines": [
{
"bbox": [
169,
82,
442,
329
],
"spans": [
{
"bbox": [
169,
82,
442,
329
],
"score": 0.976,
"type": "image",
"image_path": "ce3b790a23a18be77d8726d791255034fbb41f1cf3028b470ccdc8576c9d615a.jpg"
}
]
}
],
"index": 9.5,
"virtual_lines": [
{
"bbox": [
169,
82,
442,
94.35
],
"spans": [],
"index": 0
},
{
"bbox": [
169,
94.35,
442,
106.69999999999999
],
"spans": [],
"index": 1
},
{
"bbox": [
169,
106.69999999999999,
442,
119.04999999999998
],
"spans": [],
"index": 2
},
{
"bbox": [
169,
119.04999999999998,
442,
131.39999999999998
],
"spans": [],
"index": 3
},
{
"bbox": [
169,
131.39999999999998,
442,
143.74999999999997
],
"spans": [],
"index": 4
},
{
"bbox": [
169,
143.74999999999997,
442,
156.09999999999997
],
"spans": [],
"index": 5
},
{
"bbox": [
169,
156.09999999999997,
442,
168.44999999999996
],
"spans": [],
"index": 6
},
{
"bbox": [
169,
168.44999999999996,
442,
180.79999999999995
],
"spans": [],
"index": 7
},
{
"bbox": [
169,
180.79999999999995,
442,
193.14999999999995
],
"spans": [],
"index": 8
},
{
"bbox": [
169,
193.14999999999995,
442,
205.49999999999994
],
"spans": [],
"index": 9
},
{
"bbox": [
169,
205.49999999999994,
442,
217.84999999999994
],
"spans": [],
"index": 10
},
{
"bbox": [
169,
217.84999999999994,
442,
230.19999999999993
],
"spans": [],
"index": 11
},
{
"bbox": [
169,
230.19999999999993,
442,
242.54999999999993
],
"spans": [],
"index": 12
},
{
"bbox": [
169,
242.54999999999993,
442,
254.89999999999992
],
"spans": [],
"index": 13
},
{
"bbox": [
169,
254.89999999999992,
442,
267.24999999999994
],
"spans": [],
"index": 14
},
{
"bbox": [
169,
267.24999999999994,
442,
279.59999999999997
],
"spans": [],
"index": 15
},
{
"bbox": [
169,
279.59999999999997,
442,
291.95
],
"spans": [],
"index": 16
},
{
"bbox": [
169,
291.95,
442,
304.3
],
"spans": [],
"index": 17
},
{
"bbox": [
169,
304.3,
442,
316.65000000000003
],
"spans": [],
"index": 18
},
{
"bbox": [
169,
316.65000000000003,
442,
329.00000000000006
],
"spans": [],
"index": 19
}
]
},
{
"type": "image_caption",
"bbox": [
105,
340,
503,
361
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
339,
505,
352
],
"spans": [
{
"bbox": [
105,
339,
505,
352
],
"score": 1.0,
"content": "Figure 9: Results on BCI Competition IV-2a. ASR of PGD and UAP in Case 2), i.e., computed with head model",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
351,
361,
362
],
"spans": [
{
"bbox": [
106,
351,
361,
362
],
"score": 1.0,
"content": "(w/HM); and in Case 3), i.e., computed without head model (w/oHM).",
"type": "text"
}
],
"index": 21
}
],
"index": 20.5
}
],
"index": 15.0
},
{
"type": "image",
"bbox": [
187,
374,
424,
509
],
"blocks": [
{
"type": "image_body",
"bbox": [
187,
374,
424,
509
],
"group_id": 1,
"lines": [
{
"bbox": [
187,
374,
424,
509
],
"spans": [
{
"bbox": [
187,
374,
424,
509
],
"score": 0.97,
"type": "image",
"image_path": "24bbf9586d061f5fa7fa0a1b21aa8350f5c5b44b459c9cb84b74008f7c705ba6.jpg"
}
]
}
],
"index": 27,
"virtual_lines": [
{
"bbox": [
187,
374,
424,
386.27272727272725
],
"spans": [],
"index": 22
},
{
"bbox": [
187,
386.27272727272725,
424,
398.5454545454545
],
"spans": [],
"index": 23
},
{
"bbox": [
187,
398.5454545454545,
424,
410.81818181818176
],
"spans": [],
"index": 24
},
{
"bbox": [
187,
410.81818181818176,
424,
423.090909090909
],
"spans": [],
"index": 25
},
{
"bbox": [
187,
423.090909090909,
424,
435.36363636363626
],
"spans": [],
"index": 26
},
{
"bbox": [
187,
435.36363636363626,
424,
447.6363636363635
],
"spans": [],
"index": 27
},
{
"bbox": [
187,
447.6363636363635,
424,
459.90909090909076
],
"spans": [],
"index": 28
},
{
"bbox": [
187,
459.90909090909076,
424,
472.181818181818
],
"spans": [],
"index": 29
},
{
"bbox": [
187,
472.181818181818,
424,
484.45454545454527
],
"spans": [],
"index": 30
},
{
"bbox": [
187,
484.45454545454527,
424,
496.7272727272725
],
"spans": [],
"index": 31
},
{
"bbox": [
187,
496.7272727272725,
424,
508.9999999999998
],
"spans": [],
"index": 32
}
]
},
{
"type": "image_caption",
"bbox": [
105,
519,
504,
540
],
"group_id": 1,
"lines": [
{
"bbox": [
105,
518,
505,
530
],
"spans": [
{
"bbox": [
105,
518,
505,
530
],
"score": 1.0,
"content": "Figure 10: Power spectral density comparison of the attack with and without derivative loss term, as well as the",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
529,
230,
540
],
"spans": [
{
"bbox": [
106,
529,
230,
540
],
"score": 1.0,
"content": "original signal shown in Figure 2.",
"type": "text"
}
],
"index": 34
}
],
"index": 33.5
}
],
"index": 30.25
},
{
"type": "title",
"bbox": [
107,
559,
335,
573
],
"lines": [
{
"bbox": [
105,
558,
336,
575
],
"spans": [
{
"bbox": [
105,
558,
336,
575
],
"score": 1.0,
"content": "E CLASSIFICATION CONFUSION MATRICES",
"type": "text"
}
],
"index": 35
}
],
"index": 35
},
{
"type": "text",
"bbox": [
106,
585,
505,
673
],
"lines": [
{
"bbox": [
106,
585,
504,
597
],
"spans": [
{
"bbox": [
106,
585,
504,
597
],
"score": 1.0,
"content": "We analyze the confusion matrices before and after the proposed attack. Fig. 11 shows the confusion",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
596,
504,
607
],
"spans": [
{
"bbox": [
106,
596,
504,
607
],
"score": 1.0,
"content": "matrix of EEGNet on the Physionet dataset before the attack, where all classes can be classified with",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
607,
505,
619
],
"spans": [
{
"bbox": [
106,
607,
174,
619
],
"score": 1.0,
"content": "similar accuracy",
"type": "text"
},
{
"bbox": [
174,
607,
234,
617
],
"score": 0.88,
"content": "( 7 2 . 8 \\% - 7 3 . 5 \\% )",
"type": "inline_equation"
},
{
"bbox": [
235,
607,
505,
619
],
"score": 1.0,
"content": ". Fig. 12 shows the confusion matrices for three different propagation",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
104,
618,
506,
630
],
"spans": [
{
"bbox": [
104,
618,
153,
630
],
"score": 1.0,
"content": "parameters",
"type": "text"
},
{
"bbox": [
154,
618,
223,
630
],
"score": 0.9,
"content": "( \\lambda _ { m } \\in \\{ 1 , 5 , 1 5 \\} )",
"type": "inline_equation"
},
{
"bbox": [
223,
618,
506,
630
],
"score": 1.0,
"content": "and two attack positions (T9 and T10) which correspond to the left and",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
628,
506,
641
],
"spans": [
{
"bbox": [
105,
628,
506,
641
],
"score": 1.0,
"content": "right side of the head. When considering the attacks from the left side, shown in Fig. 12a–12c, more",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
640,
505,
652
],
"spans": [
{
"bbox": [
106,
640,
505,
652
],
"score": 1.0,
"content": "samples with ground-truth label “right” can be fooled to “rest”. This is particularly articulated in",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
651,
505,
663
],
"spans": [
{
"bbox": [
105,
651,
259,
663
],
"score": 1.0,
"content": "largely attenuated propagation model",
"type": "text"
},
{
"bbox": [
259,
651,
292,
662
],
"score": 0.86,
"content": "\\lambda _ { m } { = } 1 5 )",
"type": "inline_equation"
},
{
"bbox": [
292,
651,
505,
663
],
"score": 1.0,
"content": ". In a similar vein, attacks coming from the right side",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
661,
466,
674
],
"spans": [
{
"bbox": [
105,
661,
466,
674
],
"score": 1.0,
"content": "of the head (T10) are more effective on data with ground-truth label “left” (Fig. 12d–12f).",
"type": "text"
}
],
"index": 43
}
],
"index": 39.5
}
],
"page_idx": 15,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
26,
309,
38
],
"spans": [
{
"bbox": [
106,
26,
309,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
752,
311,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
764
],
"spans": [
{
"bbox": [
299,
750,
312,
764
],
"score": 1.0,
"content": "",
"type": "text",
"height": 14,
"width": 13
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
169,
82,
442,
329
],
"blocks": [
{
"type": "image_body",
"bbox": [
169,
82,
442,
329
],
"group_id": 0,
"lines": [
{
"bbox": [
169,
82,
442,
329
],
"spans": [
{
"bbox": [
169,
82,
442,
329
],
"score": 0.976,
"type": "image",
"image_path": "ce3b790a23a18be77d8726d791255034fbb41f1cf3028b470ccdc8576c9d615a.jpg"
}
]
}
],
"index": 9.5,
"virtual_lines": [
{
"bbox": [
169,
82,
442,
94.35
],
"spans": [],
"index": 0
},
{
"bbox": [
169,
94.35,
442,
106.69999999999999
],
"spans": [],
"index": 1
},
{
"bbox": [
169,
106.69999999999999,
442,
119.04999999999998
],
"spans": [],
"index": 2
},
{
"bbox": [
169,
119.04999999999998,
442,
131.39999999999998
],
"spans": [],
"index": 3
},
{
"bbox": [
169,
131.39999999999998,
442,
143.74999999999997
],
"spans": [],
"index": 4
},
{
"bbox": [
169,
143.74999999999997,
442,
156.09999999999997
],
"spans": [],
"index": 5
},
{
"bbox": [
169,
156.09999999999997,
442,
168.44999999999996
],
"spans": [],
"index": 6
},
{
"bbox": [
169,
168.44999999999996,
442,
180.79999999999995
],
"spans": [],
"index": 7
},
{
"bbox": [
169,
180.79999999999995,
442,
193.14999999999995
],
"spans": [],
"index": 8
},
{
"bbox": [
169,
193.14999999999995,
442,
205.49999999999994
],
"spans": [],
"index": 9
},
{
"bbox": [
169,
205.49999999999994,
442,
217.84999999999994
],
"spans": [],
"index": 10
},
{
"bbox": [
169,
217.84999999999994,
442,
230.19999999999993
],
"spans": [],
"index": 11
},
{
"bbox": [
169,
230.19999999999993,
442,
242.54999999999993
],
"spans": [],
"index": 12
},
{
"bbox": [
169,
242.54999999999993,
442,
254.89999999999992
],
"spans": [],
"index": 13
},
{
"bbox": [
169,
254.89999999999992,
442,
267.24999999999994
],
"spans": [],
"index": 14
},
{
"bbox": [
169,
267.24999999999994,
442,
279.59999999999997
],
"spans": [],
"index": 15
},
{
"bbox": [
169,
279.59999999999997,
442,
291.95
],
"spans": [],
"index": 16
},
{
"bbox": [
169,
291.95,
442,
304.3
],
"spans": [],
"index": 17
},
{
"bbox": [
169,
304.3,
442,
316.65000000000003
],
"spans": [],
"index": 18
},
{
"bbox": [
169,
316.65000000000003,
442,
329.00000000000006
],
"spans": [],
"index": 19
}
]
},
{
"type": "image_caption",
"bbox": [
105,
340,
503,
361
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
339,
505,
352
],
"spans": [
{
"bbox": [
105,
339,
505,
352
],
"score": 1.0,
"content": "Figure 9: Results on BCI Competition IV-2a. ASR of PGD and UAP in Case 2), i.e., computed with head model",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
351,
361,
362
],
"spans": [
{
"bbox": [
106,
351,
361,
362
],
"score": 1.0,
"content": "(w/HM); and in Case 3), i.e., computed without head model (w/oHM).",
"type": "text"
}
],
"index": 21
}
],
"index": 20.5
}
],
"index": 15.0
},
{
"type": "image",
"bbox": [
187,
374,
424,
509
],
"blocks": [
{
"type": "image_body",
"bbox": [
187,
374,
424,
509
],
"group_id": 1,
"lines": [
{
"bbox": [
187,
374,
424,
509
],
"spans": [
{
"bbox": [
187,
374,
424,
509
],
"score": 0.97,
"type": "image",
"image_path": "24bbf9586d061f5fa7fa0a1b21aa8350f5c5b44b459c9cb84b74008f7c705ba6.jpg"
}
]
}
],
"index": 27,
"virtual_lines": [
{
"bbox": [
187,
374,
424,
386.27272727272725
],
"spans": [],
"index": 22
},
{
"bbox": [
187,
386.27272727272725,
424,
398.5454545454545
],
"spans": [],
"index": 23
},
{
"bbox": [
187,
398.5454545454545,
424,
410.81818181818176
],
"spans": [],
"index": 24
},
{
"bbox": [
187,
410.81818181818176,
424,
423.090909090909
],
"spans": [],
"index": 25
},
{
"bbox": [
187,
423.090909090909,
424,
435.36363636363626
],
"spans": [],
"index": 26
},
{
"bbox": [
187,
435.36363636363626,
424,
447.6363636363635
],
"spans": [],
"index": 27
},
{
"bbox": [
187,
447.6363636363635,
424,
459.90909090909076
],
"spans": [],
"index": 28
},
{
"bbox": [
187,
459.90909090909076,
424,
472.181818181818
],
"spans": [],
"index": 29
},
{
"bbox": [
187,
472.181818181818,
424,
484.45454545454527
],
"spans": [],
"index": 30
},
{
"bbox": [
187,
484.45454545454527,
424,
496.7272727272725
],
"spans": [],
"index": 31
},
{
"bbox": [
187,
496.7272727272725,
424,
508.9999999999998
],
"spans": [],
"index": 32
}
]
},
{
"type": "image_caption",
"bbox": [
105,
519,
504,
540
],
"group_id": 1,
"lines": [
{
"bbox": [
105,
518,
505,
530
],
"spans": [
{
"bbox": [
105,
518,
505,
530
],
"score": 1.0,
"content": "Figure 10: Power spectral density comparison of the attack with and without derivative loss term, as well as the",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
529,
230,
540
],
"spans": [
{
"bbox": [
106,
529,
230,
540
],
"score": 1.0,
"content": "original signal shown in Figure 2.",
"type": "text"
}
],
"index": 34
}
],
"index": 33.5
}
],
"index": 30.25
},
{
"type": "title",
"bbox": [
107,
559,
335,
573
],
"lines": [
{
"bbox": [
105,
558,
336,
575
],
"spans": [
{
"bbox": [
105,
558,
336,
575
],
"score": 1.0,
"content": "E CLASSIFICATION CONFUSION MATRICES",
"type": "text"
}
],
"index": 35
}
],
"index": 35
},
{
"type": "text",
"bbox": [
106,
585,
505,
673
],
"lines": [
{
"bbox": [
106,
585,
504,
597
],
"spans": [
{
"bbox": [
106,
585,
504,
597
],
"score": 1.0,
"content": "We analyze the confusion matrices before and after the proposed attack. Fig. 11 shows the confusion",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
596,
504,
607
],
"spans": [
{
"bbox": [
106,
596,
504,
607
],
"score": 1.0,
"content": "matrix of EEGNet on the Physionet dataset before the attack, where all classes can be classified with",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
106,
607,
505,
619
],
"spans": [
{
"bbox": [
106,
607,
174,
619
],
"score": 1.0,
"content": "similar accuracy",
"type": "text"
},
{
"bbox": [
174,
607,
234,
617
],
"score": 0.88,
"content": "( 7 2 . 8 \\% - 7 3 . 5 \\% )",
"type": "inline_equation"
},
{
"bbox": [
235,
607,
505,
619
],
"score": 1.0,
"content": ". Fig. 12 shows the confusion matrices for three different propagation",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
104,
618,
506,
630
],
"spans": [
{
"bbox": [
104,
618,
153,
630
],
"score": 1.0,
"content": "parameters",
"type": "text"
},
{
"bbox": [
154,
618,
223,
630
],
"score": 0.9,
"content": "( \\lambda _ { m } \\in \\{ 1 , 5 , 1 5 \\} )",
"type": "inline_equation"
},
{
"bbox": [
223,
618,
506,
630
],
"score": 1.0,
"content": "and two attack positions (T9 and T10) which correspond to the left and",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
628,
506,
641
],
"spans": [
{
"bbox": [
105,
628,
506,
641
],
"score": 1.0,
"content": "right side of the head. When considering the attacks from the left side, shown in Fig. 12a–12c, more",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
106,
640,
505,
652
],
"spans": [
{
"bbox": [
106,
640,
505,
652
],
"score": 1.0,
"content": "samples with ground-truth label “right” can be fooled to “rest”. This is particularly articulated in",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
651,
505,
663
],
"spans": [
{
"bbox": [
105,
651,
259,
663
],
"score": 1.0,
"content": "largely attenuated propagation model",
"type": "text"
},
{
"bbox": [
259,
651,
292,
662
],
"score": 0.86,
"content": "\\lambda _ { m } { = } 1 5 )",
"type": "inline_equation"
},
{
"bbox": [
292,
651,
505,
663
],
"score": 1.0,
"content": ". In a similar vein, attacks coming from the right side",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
661,
466,
674
],
"spans": [
{
"bbox": [
105,
661,
466,
674
],
"score": 1.0,
"content": "of the head (T10) are more effective on data with ground-truth label “left” (Fig. 12d–12f).",
"type": "text"
}
],
"index": 43
}
],
"index": 39.5,
"bbox_fs": [
104,
585,
506,
674
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
239,
127,
371,
249
],
"blocks": [
{
"type": "image_body",
"bbox": [
239,
127,
371,
249
],
"group_id": 0,
"lines": [
{
"bbox": [
239,
127,
371,
249
],
"spans": [
{
"bbox": [
239,
127,
371,
249
],
"score": 0.932,
"type": "image",
"image_path": "1ea1613670bc55b7d890555d5cc7a5add225bf73d62460f527b51eaf42cf5122.jpg"
}
]
}
],
"index": 0.5,
"virtual_lines": [
{
"bbox": [
239,
127,
371,
188.0
],
"spans": [],
"index": 0
},
{
"bbox": [
239,
188.0,
371,
249.0
],
"spans": [],
"index": 1
}
]
},
{
"type": "image_caption",
"bbox": [
166,
259,
444,
270
],
"group_id": 0,
"lines": [
{
"bbox": [
165,
257,
445,
273
],
"spans": [
{
"bbox": [
165,
257,
445,
273
],
"score": 1.0,
"content": "Figure 11: Confusion matrix original EEG predictions on Physionet dataset.",
"type": "text"
}
],
"index": 2
}
],
"index": 2
}
],
"index": 1.25
},
{
"type": "image",
"bbox": [
114,
368,
491,
635
],
"blocks": [
{
"type": "image_body",
"bbox": [
114,
368,
491,
635
],
"group_id": 1,
"lines": [
{
"bbox": [
114,
368,
491,
635
],
"spans": [
{
"bbox": [
114,
368,
491,
635
],
"score": 0.974,
"type": "image",
"image_path": "f736298343192e93d83223443cacb3b79a8f52cbc4fc153ebd24626308f0133e.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
114,
368,
491,
457.0
],
"spans": [],
"index": 3
},
{
"bbox": [
114,
457.0,
491,
546.0
],
"spans": [],
"index": 4
},
{
"bbox": [
114,
546.0,
491,
635.0
],
"spans": [],
"index": 5
}
]
},
{
"type": "image_caption",
"bbox": [
106,
642,
506,
684
],
"group_id": 1,
"lines": [
{
"bbox": [
106,
643,
506,
655
],
"spans": [
{
"bbox": [
106,
643,
506,
655
],
"score": 1.0,
"content": "Figure 12: Confusion matrices for the Physionet dataset after attacking EEGNet with the proposed PGD attack",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
653,
505,
664
],
"spans": [
{
"bbox": [
105,
653,
505,
664
],
"score": 1.0,
"content": "with derivative and considering the spatial propagation. The attack is either performed from the left electrode",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
663,
506,
675
],
"spans": [
{
"bbox": [
106,
663,
467,
675
],
"score": 1.0,
"content": "(T9) or from the right electrode (T10). We consider different magnitude propagation parameters",
"type": "text"
},
{
"bbox": [
467,
663,
481,
673
],
"score": 0.91,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
481,
663,
506,
675
],
"score": 1.0,
"content": "and a",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
673,
229,
684
],
"spans": [
{
"bbox": [
105,
673,
198,
684
],
"score": 1.0,
"content": "constant delay parameter",
"type": "text"
},
{
"bbox": [
199,
673,
225,
683
],
"score": 0.88,
"content": "\\lambda _ { d } { = } 0 . 3",
"type": "inline_equation"
},
{
"bbox": [
226,
673,
229,
684
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 9
}
],
"index": 7.5
}
],
"index": 5.75
}
],
"page_idx": 16,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
26,
308,
38
],
"spans": [
{
"bbox": [
106,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
310,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
764
],
"spans": [
{
"bbox": [
299,
750,
312,
764
],
"score": 1.0,
"content": "",
"type": "text",
"height": 14,
"width": 13
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
239,
127,
371,
249
],
"blocks": [
{
"type": "image_body",
"bbox": [
239,
127,
371,
249
],
"group_id": 0,
"lines": [
{
"bbox": [
239,
127,
371,
249
],
"spans": [
{
"bbox": [
239,
127,
371,
249
],
"score": 0.932,
"type": "image",
"image_path": "1ea1613670bc55b7d890555d5cc7a5add225bf73d62460f527b51eaf42cf5122.jpg"
}
]
}
],
"index": 0.5,
"virtual_lines": [
{
"bbox": [
239,
127,
371,
188.0
],
"spans": [],
"index": 0
},
{
"bbox": [
239,
188.0,
371,
249.0
],
"spans": [],
"index": 1
}
]
},
{
"type": "image_caption",
"bbox": [
166,
259,
444,
270
],
"group_id": 0,
"lines": [
{
"bbox": [
165,
257,
445,
273
],
"spans": [
{
"bbox": [
165,
257,
445,
273
],
"score": 1.0,
"content": "Figure 11: Confusion matrix original EEG predictions on Physionet dataset.",
"type": "text"
}
],
"index": 2
}
],
"index": 2
}
],
"index": 1.25
},
{
"type": "image",
"bbox": [
114,
368,
491,
635
],
"blocks": [
{
"type": "image_body",
"bbox": [
114,
368,
491,
635
],
"group_id": 1,
"lines": [
{
"bbox": [
114,
368,
491,
635
],
"spans": [
{
"bbox": [
114,
368,
491,
635
],
"score": 0.974,
"type": "image",
"image_path": "f736298343192e93d83223443cacb3b79a8f52cbc4fc153ebd24626308f0133e.jpg"
}
]
}
],
"index": 4,
"virtual_lines": [
{
"bbox": [
114,
368,
491,
457.0
],
"spans": [],
"index": 3
},
{
"bbox": [
114,
457.0,
491,
546.0
],
"spans": [],
"index": 4
},
{
"bbox": [
114,
546.0,
491,
635.0
],
"spans": [],
"index": 5
}
]
},
{
"type": "image_caption",
"bbox": [
106,
642,
506,
684
],
"group_id": 1,
"lines": [
{
"bbox": [
106,
643,
506,
655
],
"spans": [
{
"bbox": [
106,
643,
506,
655
],
"score": 1.0,
"content": "Figure 12: Confusion matrices for the Physionet dataset after attacking EEGNet with the proposed PGD attack",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
653,
505,
664
],
"spans": [
{
"bbox": [
105,
653,
505,
664
],
"score": 1.0,
"content": "with derivative and considering the spatial propagation. The attack is either performed from the left electrode",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
663,
506,
675
],
"spans": [
{
"bbox": [
106,
663,
467,
675
],
"score": 1.0,
"content": "(T9) or from the right electrode (T10). We consider different magnitude propagation parameters",
"type": "text"
},
{
"bbox": [
467,
663,
481,
673
],
"score": 0.91,
"content": "\\lambda _ { m }",
"type": "inline_equation"
},
{
"bbox": [
481,
663,
506,
675
],
"score": 1.0,
"content": "and a",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
673,
229,
684
],
"spans": [
{
"bbox": [
105,
673,
198,
684
],
"score": 1.0,
"content": "constant delay parameter",
"type": "text"
},
{
"bbox": [
199,
673,
225,
683
],
"score": 0.88,
"content": "\\lambda _ { d } { = } 0 . 3",
"type": "inline_equation"
},
{
"bbox": [
226,
673,
229,
684
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 9
}
],
"index": 7.5
}
],
"index": 5.75
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
109,
142,
504,
632
],
"blocks": [
{
"type": "image_body",
"bbox": [
109,
142,
504,
632
],
"group_id": 0,
"lines": [
{
"bbox": [
109,
142,
504,
632
],
"spans": [
{
"bbox": [
109,
142,
504,
632
],
"score": 0.966,
"type": "image",
"image_path": "6345909357cc6d546532e1d948e55cd4d914328dd323e1190fe1921adc19c884.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
109,
142,
504,
305.33333333333337
],
"spans": [],
"index": 0
},
{
"bbox": [
109,
305.33333333333337,
504,
468.66666666666674
],
"spans": [],
"index": 1
},
{
"bbox": [
109,
468.66666666666674,
504,
632.0000000000001
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
640,
505,
662
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
640,
505,
651
],
"spans": [
{
"bbox": [
106,
640,
505,
651
],
"score": 1.0,
"content": "Figure 13: A successful PGD attack on Physionet dataset (i) without and (ii) with derivative with different",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
651,
227,
662
],
"spans": [
{
"bbox": [
106,
651,
217,
662
],
"score": 1.0,
"content": "values of maximum amplitude",
"type": "text"
},
{
"bbox": [
218,
653,
223,
660
],
"score": 0.45,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
223,
651,
227,
662
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5
}
],
"index": 2.25
}
],
"page_idx": 17,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
26,
308,
38
],
"spans": [
{
"bbox": [
106,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2022",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
311,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
763
],
"spans": [
{
"bbox": [
299,
750,
312,
763
],
"score": 1.0,
"content": "18",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
109,
142,
504,
632
],
"blocks": [
{
"type": "image_body",
"bbox": [
109,
142,
504,
632
],
"group_id": 0,
"lines": [
{
"bbox": [
109,
142,
504,
632
],
"spans": [
{
"bbox": [
109,
142,
504,
632
],
"score": 0.966,
"type": "image",
"image_path": "6345909357cc6d546532e1d948e55cd4d914328dd323e1190fe1921adc19c884.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
109,
142,
504,
305.33333333333337
],
"spans": [],
"index": 0
},
{
"bbox": [
109,
305.33333333333337,
504,
468.66666666666674
],
"spans": [],
"index": 1
},
{
"bbox": [
109,
468.66666666666674,
504,
632.0000000000001
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
640,
505,
662
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
640,
505,
651
],
"spans": [
{
"bbox": [
106,
640,
505,
651
],
"score": 1.0,
"content": "Figure 13: A successful PGD attack on Physionet dataset (i) without and (ii) with derivative with different",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
651,
227,
662
],
"spans": [
{
"bbox": [
106,
651,
217,
662
],
"score": 1.0,
"content": "values of maximum amplitude",
"type": "text"
},
{
"bbox": [
218,
653,
223,
660
],
"score": 0.45,
"content": "\\epsilon",
"type": "inline_equation"
},
{
"bbox": [
223,
651,
227,
662
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5
}
],
"index": 2.25
}
]
}
],
"_backend": "pipeline",
"_version_name": "2.2.2"
}