Hugging Face's logo

Spaces:
Z-Edgar
/
Agent-IPI-SID-Defense
Running

App Files Community
Z-Edgar commited on
Commit
ebd5196
·
verified ·
1 Parent(s): 298003d

Update index.html

Browse files
Files changed (1) hide show
  1. index.html +106 -39
index.html CHANGED
@@ -4,83 +4,150 @@
4
  <meta charset="UTF-8">
5
  <meta name="viewport" content="width=device-width, initial-scale=1.0">
6
  <title>Agent-IPI-SID-Defense</title>
 
7
  <style>
8
  body {
9
  font-family: Arial, sans-serif;
10
- text-align: center;
11
- background: linear-gradient(to right, #1E3A8A, #10B981);
12
  margin: 0;
13
- padding: 20px;
 
 
 
 
 
 
14
  color: white;
15
  }
16
  .container {
17
- max-width: 600px;
18
  margin: 0 auto;
19
- padding: 40px;
20
- background: rgba(255, 255, 255, 0.1);
21
- border-radius: 10px;
22
  }
23
  h1 {
24
- color: white;
25
- font-size: 2.5em;
26
  margin-bottom: 10px;
27
  }
28
  .subtitle {
29
- font-size: 1.2em;
30
  margin-bottom: 20px;
 
31
  }
32
  .authors {
33
  font-size: 1em;
34
  margin-bottom: 30px;
 
35
  }
36
  .buttons {
37
  margin: 30px 0;
 
 
 
38
  }
39
  button {
40
- display: inline-block;
41
- width: 100px;
42
- margin: 10px;
43
- padding: 12px 20px;
44
- font-size: 14px;
45
  border: none;
46
- border-radius: 5px;
47
  cursor: pointer;
48
- background-color: white;
49
  color: #1E3A8A;
50
- transition: background-color 0.3s;
 
 
 
 
51
  }
52
  button:hover {
53
- background-color: #f0f0f0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
54
  }
55
  .description {
56
  color: #e0e0e0;
57
  line-height: 1.6;
 
58
  }
59
  </style>
60
  </head>
61
  <body>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
62
  <div class="container">
63
- <h1>Agent-IPI-SID-Defense</h1>
64
- <p class="subtitle">NOT JUST DETECTION</p>
65
- <p class="authors"></p>
66
- <div class="buttons">
67
- <a href="#" onclick="alert('Paper link coming soon!'); return false;">
68
- <button>Paper</button>
69
- </a>
70
- <a href="https://huggingface.co/datasets/Z-Edgar/NOT_JUST_DETECTION_DATA" target="_blank">
71
- <button>Dataset</button>
72
- </a>
73
- <a href="https://huggingface.co/Z-Edgar/Agent_Safe" target="_blank">
74
- <button>Model</button>
75
- </a>
76
- <a href="https://huggingface.co/spaces/Z-Edgar/Agent_Safe" target="_blank">
77
- <button>Demo</button>
78
- </a>
 
 
 
 
 
 
 
79
  </div>
80
- <p class="description">
81
- Dataset: 400k+ prompt injection examples for training safe agents.<br>
82
- Model: Agent_Safe.
83
- </p>
84
  </div>
85
  </body>
86
  </html>
 
4
  <meta charset="UTF-8">
5
  <meta name="viewport" content="width=device-width, initial-scale=1.0">
6
  <title>Agent-IPI-SID-Defense</title>
7
+ <link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.2/css/all.min.css">
8
  <style>
9
  body {
10
  font-family: Arial, sans-serif;
 
 
11
  margin: 0;
12
+ color: #333;
13
+ background-color: #f4f4f4;
14
+ }
15
+ .header {
16
+ background: linear-gradient(to right, #1E3A8A, #10B981);
17
+ text-align: center;
18
+ padding: 60px 20px;
19
  color: white;
20
  }
21
  .container {
22
+ max-width: 800px;
23
  margin: 0 auto;
24
+ padding: 20px;
 
 
25
  }
26
  h1 {
27
+ font-size: 2.8em;
 
28
  margin-bottom: 10px;
29
  }
30
  .subtitle {
31
+ font-size: 1.3em;
32
  margin-bottom: 20px;
33
+ opacity: 0.9;
34
  }
35
  .authors {
36
  font-size: 1em;
37
  margin-bottom: 30px;
38
+ opacity: 0.8;
39
  }
40
  .buttons {
41
  margin: 30px 0;
42
+ display: flex;
43
+ justify-content: center;
44
+ gap: 15px;
45
  }
46
  button {
47
+ display: inline-flex;
48
+ align-items: center;
49
+ padding: 12px 24px;
50
+ font-size: 16px;
51
+ font-weight: 500;
52
  border: none;
53
+ border-radius: 25px;
54
  cursor: pointer;
55
+ background-color: #ffffff;
56
  color: #1E3A8A;
57
+ box-shadow: 0 4px 6px rgba(0, 0, 0, 0.1);
58
+ transition: all 0.3s ease;
59
+ }
60
+ button i {
61
+ margin-right: 8px;
62
  }
63
  button:hover {
64
+ background-color: #10B981;
65
+ color: white;
66
+ transform: translateY(-2px);
67
+ box-shadow: 0 6px 12px rgba(0, 0, 0, 0.15);
68
+ }
69
+ .content {
70
+ background: white;
71
+ padding: 40px;
72
+ border-radius: 10px;
73
+ box-shadow: 0 2px 10px rgba(0, 0, 0, 0.1);
74
+ margin-top: -20px;
75
+ }
76
+ .abstract {
77
+ line-height: 1.6;
78
+ color: #444;
79
+ margin-bottom: 30px;
80
+ }
81
+ .images {
82
+ display: flex;
83
+ justify-content: center;
84
+ gap: 20px;
85
+ flex-wrap: wrap;
86
+ }
87
+ .images img {
88
+ max-width: 100%;
89
+ height: auto;
90
+ border-radius: 8px;
91
+ box-shadow: 0 2px 5px rgba(0, 0, 0, 0.1);
92
  }
93
  .description {
94
  color: #e0e0e0;
95
  line-height: 1.6;
96
+ margin-top: 20px;
97
  }
98
  </style>
99
  </head>
100
  <body>
101
+ <div class="header">
102
+ <div class="container">
103
+ <h1>Agent-IPI-SID-Defense</h1>
104
+ <p class="subtitle">NOT JUST DETECTION</p>
105
+ <p class="authors"></p>
106
+ <div class="buttons">
107
+ <a href="#" onclick="alert('Paper link coming soon!'); return false;">
108
+ <button><i class="fas fa-file-alt"></i> Paper</button>
109
+ </a>
110
+ <a href="https://huggingface.co/datasets/Z-Edgar/NOT_JUST_DETECTION_DATA" target="_blank">
111
+ <button><i class="fas fa-database"></i> Dataset</button>
112
+ </a>
113
+ <a href="https://huggingface.co/Z-Edgar/Agent_Safe" target="_blank">
114
+ <button><i class="fas fa-cogs"></i> Model</button>
115
+ </a>
116
+ <a href="https://huggingface.co/spaces/Z-Edgar/Agent_Safe" target="_blank">
117
+ <button><i class="fas fa-eye"></i> Demo</button>
118
+ </a>
119
+ </div>
120
+ <p class="description">
121
+ Dataset: 400k+ prompt injection examples for training safe agents.<br>
122
+ Model: Agent_Safe.
123
+ </p>
124
+ </div>
125
+ </div>
126
  <div class="container">
127
+ <div class="content">
128
+ <h2>Abstract</h2>
129
+ <p class="abstract">
130
+ This paper proposes an innovative defense framework against indirect prompt injection attacks in agent systems, integrating structureaware attention-based detection with preference-aligned purification. Our method effectively identifies and removes malicious
131
+ instructions embedded within structured interaction data, while
132
+ preserving task utility and ensuring model security. Specifically,
133
+ we design an end-to-end structured defense pipeline that combines
134
+ supervised fine-tuning and reinforcement-based policy optimization to accurately filter adversarial content without compromising
135
+ structural integrity. To support the training of the purification module, we construct the first adversarial dataset tailored to structured
136
+ indirect injection scenarios. Furthermore, we introduce a novel
137
+ attack variant that manipulates response data fields to simulate
138
+ more deceptive and realistic threats targeting agent behavior. Experimental results on the AgentDojo benchmark demonstrate that,
139
+ compared to existing detection-based defenses, our method not
140
+ only significantly reduces attack success rates but also substantially
141
+ improves the agent’s task completion performance in interactive
142
+ settings.
143
+ </p>
144
+ <h2>Figures</h2>
145
+ <div class="images">
146
+ <!-- Placeholder images; replace with actual image URLs -->
147
+ <img src="https://via.placeholder.com/300x200?text=Figure+1" alt="Figure 1">
148
+ <img src="https://via.placeholder.com/300x200?text=Figure+2" alt="Figure 2">
149
+ </div>
150
  </div>
 
 
 
 
151
  </div>
152
  </body>
153
  </html>