| Attacks | Method | OS | Year | Strategy | Setting | Datasets | Target Models | Metrics |
| Model Extraction | Carlini et al. [503] | Yes | 2024 | Binary Search | Black-box | None | GPTs, LLaMA, Pythia,ada, cabbage | Query&TokenCost, MSE, RMS |
| Finlayson et al. [504] | No | 2024 | Softmax Bottleneck | Black-box | None | Pythia, GPT-3.5 | Query Cost |
| Zanella et al. [505] | No | 2024 | Matrix Operations | Grey-box | SST-2, MNLI, AGNews | BERTs, XLNet | Query Cost, Acc,Agreement |
| Horwitz et al. [506] | Yes | 2024 | Spectral DeTuning | White-box | LoWRA | ViT, SD, Mistral | MSWE, SEM |
| Membership Inference | MIN-K% PROB [509] | Yes | 2023 | Probabilities | Black-box | Wikipedia | LLaMAs, Pythia, NeoX,OPT | TPR, FPR, ROC,AUC |
| MIN-K%++ [510] | Yes | 2022 | Local Maxima | Black-box | WikiMIA, MIMIR | Pythia, GPT-NeoXLLaMA, OPT, Mamba | AUROC, TPR, FPR |
| Blind [511] | Yes | 2024 | Threshold | Black-box | 8 sets | GPT-3, OpenLLaMA | AUC ROC |
| LLM-DI [512] | Yes | 2024 | Aggregation | Black-box | PILE | Pythias | AUC, p-values |
| DE-COP [513] | Yes | 2024 | Paraphrases | Black-box | arXiv:Tection, BookTection | Mistral, Mixtral, LLaMA, GPTs, Claude | AUC |
| Recall [514] | Yes | 2024 | Log-Likelihoods | Black-box | WikiMIA, MIMIR | Pythia, GPT-NeoXLLaMA, OPT, Mamba | AUC, TPR@FPR |
| Noisy [515] | No | 2024 | Embedding NGBRs | Gray-box | OpenWebText,Wikipedia | GPT-2 | TPR, FPR, AUC |
| SMIA [516] | No | 2024 | Perturbation | Gray-box | Wikipedia, FAN | Pythia, Pythia-Deduped, GPT-Neos | AUC-ROC, TPR, FPR |
| FEATAGG [517] | No | 2024 | Feature Aggregation | Black-box | ProjectGutenberg,ArXiv | OpenLLaMA | TPR@FPR, AUC |
| RAG-MIA [520] | No | 2024 | Direct Asking | Black-box | HealthCareMagic,Enron | flan, llama, mistral | TPR@FPR, AUC-ROC |
| Jailbreak | GCG [260] | Yes | 2023 | Gradient-based | White-box | Vicuna, LLaMA-2 | AdvBench | ASR, Loss |
| AmpleGCG [530] | Yes | 2024 | Hybrid-based | White-box | Vicuna, Llama-2,Mis-tral,GPTs | AdvBench | ASR, US, Diver-sity, Time |
| I-GCG [531] | Yes | 2024 | Gradient-based | White-box | AdvBench,HarmBench | VICUNA, GUANACOLLAMA, MISTRAL | ASR |
| MA-GCG [532] | Yes | 2024 | Gradient-based | White-box | AdvBench | Vicuna, Mistral | ASR, Time |
| A-GCG [533] | Yes | 2024 | Gradient-based | White-box | AdvBench | Llama2, Vicuna | ASR, Acc |
| AutoDAN-A [534] | Yes | 2023 | LLM-based | Black-box | AdvBench | Vicuna, Mistral | ASR, Recheck,PPL |
| AutoDAN-B [535] | Yes | 2023 | Gradient-based | White-box | AdvBench | Vicuna, Guanaco, Pythia | ASR, Recheck |
| PAIR [261] | Yes | 2023 | LLM-based | Black-box | JailbreakBench | Vicuna, Llama-2, GPTs,Claudes,Gemini | ASR, QPS |
| ToA [536] | Yes | 2023 | LLM-based | Black-box | AdvBench, Harm123 | Vicuna, Llama-2, PaLM-2,GPTs, Claude3, Gemini | GPT4-MetricHuman-Judge |
| PAL [537] | Yes | 2024 | LLM-based | Black-box | AdvBench | Llama-2, GPT-3.5 | ASR, Manual Labeling |
| Masterkey [538] | No | 2023 | Rephrasing | Black-box | AdvBench, Harm123 | GPTs, Bing, Bard | ASR, QSR |
| AutoDAN-Turbo [539] | Yes | 2024 | LLM-based | Black-box | Harmbench | Llama-2, Gemma, GPT-4,Gemini | ASR, StrongRE-JECT |
| FlipAttack [540] | Yes | 2025 | Rephrasing | Black-box | AdvBench, StrongRE-JECT | GPTs, Claude 3.5 Sonnet, Llama 3.1 405B, Mixtral 8x22B | ASR |
| Geneshift [541] | Yes | 2025 | LLM-based | Black-box | AdvBench | GPTs | ASR |
| Prompt Injection | IPP [542] | Yes | 2022 | Handcraft | Black-box | OpenAI Examples | text-davinci | ASR |
| Greshake et al. [543] | Yes | 2023 | Data Poisoning | Black-box | None | text-davinci, GPT-4 | None |
| HOUYI [544] | Yes | 2023 | Components Asmbl | Black-box | Five Queries | SUPERTOOLS | Manual |
| Yan et al. [130] | Yes | 2023 | Poisoning | Black-box | Several Cases | Alpaca | Ngt, Pst, Ocrc |
| TT [545] | No | 2023 | Game | Black-box | Tensor Trust | GPTs, Claudes, PaLM, LLaMAs | Robustness Rate |
| JudgeDeceiver [546] | Yes | 2024 | Gradient-based | White-box | MT-Bench, LLMBarMRPC, Jfleg, HSOL,RTE, SST2, SMS | Mistral, Openchat, Llamas | ACC, ASR, PACKEY-E, LM-E |
| AUPI [547] | Yes | 2024 | Gradient-based | White-box | MRPC, Jfleg, HSOL,RTE, SST2, SMS | Llama2 | ASR |
| AUTOHIJACKER [548] | No | 2024 | LLM-based | Black-box | AgentDojo, OPI | Llama, Command-R,GPTs | ASR |
| Data Extraction | zlib [108] | Yes | 2020 | Generate & Inference | Black-box | Top-n, Temperature, Internet | GPT-2 | 6 metrics |
| AutoSklearn [549] | No | 2023 | Greedy, Contrastive, Beam decoding | Black-box | Pile | GPT-Neo | Precision, Recall,R@FPR |
| DECOM [550] | No | 2024 | Decomposition | Black-box | NYT, WSJ | Frontiers | TRM, EMP,BITAP |
| Context [551] | No | 2022 | Context, Zero-shot,Few-shot | Black-box | Enron Corpus | GPT-Neo | Acc |
| ETHICIST [552] | Yes | 2023 | Prompt Tuning | Gray-box | LM-Extraction | GPT-Neo | Recall |
| Pli-compass [553] | No | 2024 | Grounding | Black-box | Enron email | GPT-J | Extraction Rate |
| DSP [554] | No | 2024 | Dynamic Soft Prompting | Black-box | LMEB, The Stack | GPT-Neo, Pythia, Star-CoderBase | EER, FER, PPL |
| PWB [555] | Yes | 2024 | Gradient-based | White-box | Pile | Pythia, Llama | Precision, AUC,TPR |
| Prompt Stealing | Sha et al. [556] | No | 2024 | LLM-based | Black-box | RetrievalQA,AlpacaGPT4 | ChatGPT, LLaMA | Acc, Precision, Recall, AUC |
| output2prompt [557] | Yes | 2024 | LLM-based | Black-box | 3 User & 3 SystemPrompts | Llamas, GPTs | BLEU, CS, Preci-sion, Recall |
| PRSA [558] | No | 2024 | Output Difference | Black-box | Category18 | GPTs | BLEU, FastKAS-SIM, JS |
",
+ "bbox": [
+ 76,
+ 126,
+ 921,
+ 902
+ ],
+ "page_idx": 19
+ },
+ {
+ "type": "header",
+ "text": "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015",
+ "bbox": [
+ 73,
+ 31,
+ 421,
+ 44
+ ],
+ "page_idx": 19
+ },
+ {
+ "type": "page_number",
+ "text": "20",
+ "bbox": [
+ 906,
+ 32,
+ 921,
+ 42
+ ],
+ "page_idx": 19
+ },
+ {
+ "type": "text",
+ "text": "ECLIPSE [588] uses an LLM as a suffix generator and optimizer. PAL [537] proposes an online proxy model (which is used for adversarial prompt generation) training pipeline.",
+ "bbox": [
+ 84,
+ 51,
+ 490,
+ 111
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "$\\Rightarrow$ Others. EnJa [589] proposes to ensemble prompt and token-level attack methods via a template-based connector. AmpleGCG [530] first collects lots of successful suffixes and then trains the generative model to generate a specific suffix for a given unsafe prompt. Zhao et al. [590] targets the scenario where the decoding process of target LLM is assisted with smaller models' guidance.",
+ "bbox": [
+ 66,
+ 111,
+ 491,
+ 214
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "Prompt Injection Attacks. Prompt injection is a vulnerability where an attacker manipulates the input prompts of LLMs to force them to generate a specific output, which is usually out of the range for normal use (e.g., goal hijacking and prompt leaking [542]), often by injecting malicious text or commands into the input field. Attackers can employ a variety of techniques to carry out such attacks.",
+ "bbox": [
+ 66,
+ 220,
+ 491,
+ 323
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "$\\Rightarrow$ Direct Prompt Injection. Perez et al. [542] directly inject handcrafted adversarial prompts into inputs to misalign the language model. HOUYI [544] proposes an injection generation framework which includes three components. Yan et al. [130] utilize LLMs to generate diverse trigger instructions that implicitly capture the characteristics of trigger scenarios. TENSOR TRUST leverages the TENSOR TRUST web game to generate a large-scale dataset and benchmark [545]. AUPI [547] adopts a gradient-based optimization method, specifically, a momentum-enhanced optimization algorithm, to generate universal prompt injection data. Upadhayay et al. [591] argue that LLMs suffer from cognitive overload and propose to use in-context learning to jailbreak LLMs through deliberately designed prompts that induce cognitive overload. Kwon et al. [592] circumvent security policies by substituting sensitive words—likely to be rejected by the language model—with mathematical functions.",
+ "$\\nLeftrightarrow$ Indirect Prompt Injection. Greshake et al. [543] propose to indirectly inject prompts into the data that are likely to be retrieved. Bagdasaryan et al. [593] design a prompt injection attack against multi-modal LLMs, by generating an adversarial perturbation corresponding to the prompt and blending it into an image or audio recording. Neural Exec [594] designs a multi-stage preprocessing pipeline for cases like Retrieval-Augmented Generation (RAG)-based applications. PoisonedAlign [595] boosts the success of prompt injection attacks by strategically creating poisoned alignment samples in the LLM's alignment process. TPIA [596] crafts non-functional perturbations that contain malicious information and inserts them into the victim's code context by spreading them into potentially used dependencies like packages or RAG's knowledge base. F2A [597] proposes to use feign security detection agents to bypass the defense mechanism of LLMs. AUTOHIJACKER [548] uses a batch-based optimization framework to handle sparse feedback and leverages a trainable memory to enable effective generation.",
+ "Different Settings. JudgeDeceiver uses gradient-based optimization to inject LLM-as-a-Judge scenarios [546]. Pedro et al. [598] study the risk of injections targeting web applications based on the Langchain framework. Lee et"
+ ],
+ "bbox": [
+ 66,
+ 329,
+ 491,
+ 943
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "al. [599] propose a human-AI collaborative framework to explore the potential of prompt injection against federated military LLMs. PROMPT INFECTION [600] proposes to make malicious prompts self-replicate across interconnected agents in multi-agent systems. Zhang et al. [601] explore the risk of prompt injection in LLM-integrated systems like LLM-integrated mobile robotic systems.",
+ "bbox": [
+ 514,
+ 51,
+ 921,
+ 156
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "Data Extraction Attacks. Data extraction attacks try to figure out the personally identifiable information (PII) that is used to train the LLMs [108]. It starts from sufficient-length prefixes to perform extraction and additional measures to determine if extracted texts are valid.",
+ "bbox": [
+ 503,
+ 162,
+ 923,
+ 234
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "$\\nLeftrightarrow$ Methods. In the beginning work [108], the proposed extraction process contains two stages \"generate-then-rank\": sampling potentially memorized examples and membership inference. It proposes a temperature-decaying method to sample more diverse examples and use surrogate models to infer the membership. After that, Al-Kaswan et al. [549] propose using greedy, contrastive, and beam decoding strategies to generate examples and use a classifier to infer the membership. Su et al. [550] propose an instruction decomposition technique to extract fragments of training data gradually. Huang et al. [551] extensively explore the effect of context, zero-shot, and few-shot methods in extracting the personal email address. ETHICIST proposes a smoothing loss and a calibrated confidence estimation method to extract the suffix and measure the confidence [552]. Nakka et al. [553] improves the extraction performance by grounding the prefix of the manually constructed extraction prompt with in-domain data. Wang et al. [554] propose to train a transformer-based generator to produce dynamic, prefix-dependent soft prompts. Ozdayi et al. [105] introduce an approach that uses prompt tuning to control the extraction rates of memorized content. Meng et al. [602] propose a two-stage method, i.e., collection and ranking, to recover PPI when PII entities have been masked.",
+ "Different Settings. Some works also explore the risk of data leakage in novel settings. Wang et al. [555] study the probability of data extraction in fine-tuning settings and Bargav et al. [603, 604] extract the training data by comparing the output difference before and after the fine-tuning. Jiang et al. [605, 606, 607] propose to extract the private Retrieval-Augmented Generation (RAG) documents. Peng et al. [608] extract the private RAG documents by poisoning in the fine-tuning process. Nasr et al. [107] explore the potential risk of data extraction for the aligned production language models. Panda et al. [609] extract the fine-tuning secret data by poisoning the pertaining dataset. Lu et al. [610] propose to extract PII from an aligned model with model merging. Chen et al. [611] find that fine-tuning can recover the forgotten PIIs in pretraining data. Panchendrarajan et al. [612] propose to extract the whole private training data in the fine-tuning process. Rashid et al. [613] propose selective weight tampering to explore PPI leakage in Federated Language Models. Dentan et al. [614] extract data from layout-aware document understanding models like unimodal or bimodal models.",
+ "Different Applications. Leveraging the abnormally high"
+ ],
+ "bbox": [
+ 500,
+ 242,
+ 923,
+ 943
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "header",
+ "text": "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015",
+ "bbox": [
+ 73,
+ 31,
+ 421,
+ 44
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "page_number",
+ "text": "21",
+ "bbox": [
+ 906,
+ 32,
+ 919,
+ 42
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "text",
+ "text": "token probabilities, some works utilize the memorization of LLMs to extract the fingerprint or steganography [615]. Al-Kaswan et al. [616] explore memorization in large language models for code and find that code models memorize training data at a lower rate than natural language models. Nie et al. [617] utilize the token-level features derived from the identified characteristics to decode the PII. Lehman et al. [618] reveal the risk of Electronic Health Records leakage of LLMs. Diera et al. [619] conduct experiments to assess the PII leakage of fine-tuned BERT models and found that Differential Privacy (DP) has a negative effect when deployed in fine-tuning. Zhang et al. [620] propose data extraction attacks against text classification with transformers. Huang et al. [621] propose an evaluation tool, i.e. HCR, to assess the PPI leakage in Neural Code Completion Tools.",
+ "bbox": [
+ 84,
+ 53,
+ 491,
+ 286
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "text",
+ "text": "$\\nrightarrow$ Factor Assessment. Some work studies the factors of data extraction including decoding schemes, model sizes, prefix lengths, partial sequence leakages, and token positions [622, 623]. Yash et al. [624] explore the effects of prompt sensitivity and access to multiple checkpoints to extraction attacks. Staab et al. [625] construct a dataset consisting of real Reddit profiles to extract personal attributes. Xu et al. [626] conduct experiments to evaluate the factors of different suffix generation methods and different membership inference attacks in extraction performance. Karamolegkou et al. [627] evaluate the effect of model structure, data type, probing strategies, and metrics.",
+ "bbox": [
+ 68,
+ 287,
+ 491,
+ 474
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "text",
+ "text": "Prompt Stealing Attacks. Given that crafting effective prompts requires significant engineering effort and can be considered valuable intellectual property (IP), promptstealing attacks aim to compromise this IP by reconstructing prompts from generated responses [556, 557, 558]. These generation effects are often used to attract prospective prospective buyers. Sha et al. [556] pioneer this approach by collecting a dataset and training classifiers to predict prompt parameters—such as whether the prompt is direct, role-based, or in-context. They then used a large language model (LLM) to reconstruct the prompt. Similarly, Zhang et al. [557] trained an LLM on output-prompt pairs to directly infer the original prompt, while Yang et al. [558] leveraged generation differences to refine surrogate prompts. However, recovering the original prompt solely from the output is challenging. Out of this, Zheng et al. [628] propose a timing-based side-channel method to infer the prompt during inference.",
+ "bbox": [
+ 71,
+ 477,
+ 491,
+ 739
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "text",
+ "text": "6.1.2 Defensive Mechanisms in Deployment",
+ "text_level": 1,
+ "bbox": [
+ 73,
+ 750,
+ 393,
+ 765
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "text",
+ "text": "In Subsubsection 6.1.1, we analyzed various attack scenarios targeting individual LLM deployments. However, in real-world applications, defense mechanisms are not designed as isolated, one-to-one countermeasures against specific attacks. Instead, they follow fundamental security principles to establish a systematic defense framework, as illustrated in Figure 9. This framework integrates multiple layers of protection, ensuring resilience against a wide range of adversarial threats while maintaining model usability and efficiency.",
+ "bbox": [
+ 71,
+ 767,
+ 490,
+ 912
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "text",
+ "text": "Input Preprocessing Defenses Input preprocessing serves as the first line of defense in LLM deployment, aiming to",
+ "bbox": [
+ 71,
+ 912,
+ 491,
+ 944
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "image",
+ "img_path": "images/f4ce229cb80a8c96656ded2aa655c39cbe884dd49ac1e9e0e1650264348f5451.jpg",
+ "image_caption": [
+ "Defensive Mechanisms in Deployment",
+ "Fig. 9: The overview of attacks in single LLM's deployment phase."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 506,
+ 71,
+ 915,
+ 159
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "text",
+ "text": "detect and neutralize adversarial inputs before they reach the model.",
+ "bbox": [
+ 503,
+ 224,
+ 921,
+ 253
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "text",
+ "text": "Attack Detection & Identification: Effective input filtering [629, 630] begins with attack detection [631], which identifies adversarial prompts through statistical [632], structural [633], or behavioral inconsistencies [634]. Gradient-based detection methods [635] leverage safety-critical gradient analysis and loss landscape exploration to uncover jailbreak prompts that manipulate LLM behavior. These approaches identify adversarial inputs [636, 637] by analyzing how small perturbations [638] affect model outputs, detecting highly sensitive or misaligned gradients that indicate targeted attacks. Perplexity-based methods [632, 632] measure the probability distribution of input sequences, flagging atypical or low-likelihood prompts as potential adversarial inputs. These techniques are particularly effective in detecting prompt injection and adversarial perturbations, where crafted prompts deviate significantly from natural language distributions.",
+ "bbox": [
+ 501,
+ 255,
+ 921,
+ 503
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "text",
+ "text": "Beyond individual heuristics, universal detection frameworks [639] integrate multiple detection strategies to counter diverse attack vectors, including prompt injection [640], backdoor manipulations [641], and adversarial attacks [637]. These frameworks employ ensemble-based filtering mechanisms, combining gradient analysis [642], perplexity estimation [643], and syntactic evaluation for generalized attack resilience.",
+ "bbox": [
+ 501,
+ 503,
+ 921,
+ 619
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "text",
+ "text": "Semantic & Behavioral Analysis: Attack detection alone is insufficient, as certain adversarial inputs may bypass traditional filtering mechanisms. Semantic [644] and behavioral analysis enhance input preprocessing by evaluating linguistic intent and model alignment. Self-examination techniques allow LLMs [645, 646] to assess whether they are being manipulated, leveraging auxiliary reasoning steps to detect deceptive prompts. Alignment-based verification [647] ensures that the model's responses remain consistent with its safety objectives [330], identifying inputs that subtly nudge the model toward policy violations or ethical misalignment. Intention analysis [648, 649] further refines input filtering by discerning subtle manipulations designed to bypass explicit security checks. Unlike token-level detection, which flags overtly adversarial inputs, intention-aware defenses analyze the semantic structure and purpose of the input to preemptively reject jailbreak attempts.",
+ "bbox": [
+ 501,
+ 621,
+ 921,
+ 869
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "text",
+ "text": "Adversarial Defense & Mitigation: When detection and behavioral analysis fail to fully neutralize adversarial inputs, robustness-enhancing techniques [647] mitigate their effects by reducing model susceptibility to manipulation [334, 650]. Semantic smoothing [651, 652] techniques",
+ "bbox": [
+ 503,
+ 869,
+ 921,
+ 944
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "header",
+ "text": "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015",
+ "bbox": [
+ 73,
+ 32,
+ 421,
+ 44
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "page_number",
+ "text": "22",
+ "bbox": [
+ 906,
+ 32,
+ 921,
+ 42
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "text",
+ "text": "introduce controlled randomness into LLM responses, reducing the model's sensitivity to adversarial perturbations and preventing reliable jailbreak execution. By stabilizing decision boundaries [653], these methods enhance resistance against prompt manipulation strategies that exploit response predictability.",
+ "bbox": [
+ 71,
+ 53,
+ 491,
+ 141
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "Preemptive input transformations [654], such as back-translation [655] or paraphrasing, modify incoming queries [651] while preserving semantic intent, disrupting adversarial structures embedded within malicious prompts. Data augmentation [656] and adversarial training further strengthen model robustness by exposing LLMs to adversarial prompts during training, forcing them to learn invariances that reduce their vulnerability to real-world attacks.",
+ "bbox": [
+ 71,
+ 141,
+ 491,
+ 257
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "Output Filtering Mechanisms. Output filtering mechanisms [212, 657] serve as a critical safeguard in LLM deployment, ensuring that generated responses comply with safety constraints while preserving informativeness. Unlike input preprocessing, which aims to prevent adversarial prompts from reaching the model, output filtering mitigates harmful content post-generation. Existing approaches primarily follow three paradigms: rule-based constraints, generative adversarial filtering, and toxicity detection.",
+ "bbox": [
+ 71,
+ 257,
+ 491,
+ 388
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "Rule-based mechanisms [658] impose predefined constraints on model outputs, preventing the generation of harmful, unethical, or undesired content. Programmable guardrails [659] offer a structured framework where developers can enforce response filtering, topic restriction, and ethical alignment. These methods often integrate reinforcement learning from human feedback [155] or rule-based reward [660] modeling to refine output safety. While effective at handling explicit violations, static rule-based methods struggle with nuanced adversarial prompts and subtle misalignments.",
+ "bbox": [
+ 76,
+ 388,
+ 491,
+ 547
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "To address these limitations, generative adversarial filtering [661] leverages self-critique [662, 663], ensemble detection, and dynamic response evaluation [664]. Self-rectification mechanisms [663, 665] enable LLMs to critique their own outputs and refine responses through iterative refinement. Additionally, ensemble-based [666] moderation models aggregate predictions from multiple LLMs, improving robustness against circumvention techniques. Adaptive filtering frameworks [667] employ perplexity-based assessments and adversarial perturbation detection to flag responses deviating from expected linguistic patterns, enhancing their resilience against jailbreak attempts [668, 669] and toxic content injection.",
+ "bbox": [
+ 71,
+ 549,
+ 491,
+ 737
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "Toxicity detection [670, 671, 672] and content moderation [673, 674, 675, 676] further reinforce output safety by identifying and mitigating hate speech [677], misinformation, and other harmful content. Supervised finetuning adapts LLMs to recognize undesirable patterns, while classifier-based detection models [678] filter responses in real-time. Some approaches introduce debiasing strategies, such as controlled decoding [679, 680] and anti-expert guidance [681], to suppress toxic outputs without sacrificing response diversity. However, these methods face challenges in balancing false positives and false negatives, particularly in ambiguous or context-dependent cases.",
+ "bbox": [
+ 71,
+ 738,
+ 491,
+ 912
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "The effectiveness of output filtering hinges on its ability to balance strict control with linguistic flexibility, ensur",
+ "bbox": [
+ 73,
+ 912,
+ 491,
+ 941
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "ing that models remain both safe and practically useful. A hybrid approach combining rule-based safeguards, self-correcting mechanisms, and adaptive toxicity moderation is essential to achieving robust and scalable LLM deployment. Robust Prompt Engineering. Robust prompt engineering aims to enhance LLM safety by designing input prompts that resist adversarial manipulation [682], protect sensitive data, and mitigate harmful outputs—all [683] without modifying model parameters. These strategies act at the interaction level, offering lightweight and model-agnostic protection.",
+ "bbox": [
+ 501,
+ 53,
+ 921,
+ 212
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "Recent efforts have introduced prompt optimization techniques grounded in adversarial robustness, including embedding-space manipulation and defensive objective alignment. Methods such as Robust Prompt Optimization [684] and Prompt Adversarial Tuning generate transferable suffixes [668] or prefix [685] embeddings to guide model behavior [686] under attack [687], effectively lowering jailbreak success rates while preserving task performance. Similarly, goal prioritization frameworks [688] enforce inference-time objective consistency, dynamically resolving conflict between user instructions and safety constraints without requiring access to malicious samples. Complementary to these strategies, patch-based methods integrate interpretable suffixes or structured self-reminders [689] into prompts, reducing the model's susceptibility to coercive inputs through lightweight, modular defenses.",
+ "bbox": [
+ 503,
+ 213,
+ 923,
+ 460
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "Structural manipulation approaches [690] neutralize adversarial intent through prompt rewriting. Spotlighting [691] injects source-attribute signals to counter indirect prompt injection, while inverse prompt engineering [692] repurposes attack data to generate task-specific defensive prompts under the principle of least privilege.",
+ "bbox": [
+ 503,
+ 460,
+ 923,
+ 549
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "Privacy-preserving prompt [693] design introduces formal guarantees through differential privacy. Approaches like DP-Prompt [694] and stochastic gradient masking [695] reduce information leakage from prompts without harming performance. Desensitization and directional control of incontext representations offer additional privacy protections during prompt construction. Prompt engineering [579, 696] also helps mitigate societal risks. Chain-of-thought prompting and guided templates reduce gender bias [697] in reasoning tasks, while prompt learning [698] improves toxicity detection and generation control [699, 700], often surpassing specialized models in efficiency and generalization.",
+ "bbox": [
+ 503,
+ 549,
+ 923,
+ 723
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "Finally, systematic prompt optimization methods [701, 702] aim to generalize prompt robustness across tasks and domains. Techniques like BATPrompt [703] and StraGo [704] use adversarial simulation and strategic decomposition to refine prompts iteratively, improving both resilience and effectiveness under variable inputs.",
+ "bbox": [
+ 503,
+ 723,
+ 921,
+ 811
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "System-level Security Controls. System-level defenses [705] enhance LLM deployment by optimizing inference, enforcing alignment, isolating untrusted inputs, and securing the supply chain. Systems like Petals [706], Sarathi-Serve [707], and DistServe [708] restructure computation to improve throughput and latency, while TriForce [709], Medusa [710] MagicDec [711] accelerate generation via speculative decoding and structural compression. Parallel frameworks such as DeepSpeed-FastGen [712] and SpecExec [713] further boost",
+ "bbox": [
+ 503,
+ 811,
+ 923,
+ 941
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "header",
+ "text": "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015",
+ "bbox": [
+ 73,
+ 31,
+ 421,
+ 44
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "page_number",
+ "text": "23",
+ "bbox": [
+ 906,
+ 32,
+ 921,
+ 42
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "text",
+ "text": "efficiency with minimal overhead.",
+ "bbox": [
+ 71,
+ 53,
+ 313,
+ 66
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "text",
+ "text": "Runtime alignment methods [714] adapt model behavior through cross-model guidance or token-level reward modeling. Systems such as SelfDefend [715] and Gradient Cuff [716] detect unsafe generation by monitoring agreement across models or loss landscapes, while Spotlighting [691] inserts provenance signals to mitigate indirect prompt injection.",
+ "bbox": [
+ 71,
+ 66,
+ 490,
+ 170
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "text",
+ "text": "Access isolation is achieved through policy enforcement [717] and system wrappers [688]. At the supply level, tools like MalHug [718] identify poisoned models, while system audits reveal sandbox and plugin vulnerabilities, highlighting the need for end-to-end secure deployment.",
+ "bbox": [
+ 71,
+ 170,
+ 491,
+ 244
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "text",
+ "text": "LLM-based guard models utilize lightweight LLMs like Llama Guard [330], Aegis Guard [719, 720], WildGuard [721], and ShieldGemma [722] to moderate both the input and output of the victim LLMs. However, they are purely classifiers. To solve this problem, the first reasoning-based guard model named GuardReasoner [723] is proposed to improve the performance, explainability, and generalization ability via learning to reason. It brings new opportunities for the safety of large-scale reasoning models [724].",
+ "bbox": [
+ 71,
+ 244,
+ 490,
+ 377
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "text",
+ "text": "6.1.3 Evaluation and Benchmarks in Deployment",
+ "text_level": 1,
+ "bbox": [
+ 73,
+ 392,
+ 426,
+ 407
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "text",
+ "text": "To assess the reliability and safety of LLMs after deployment, evaluation efforts focus on several key dimensions and risk types, as illustrated in Figure 10. These dimensions guide the design of systematic benchmarks and metrics tailored for real-world deployment settings.",
+ "bbox": [
+ 71,
+ 411,
+ 491,
+ 486
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "image",
+ "img_path": "images/7e5d6796694a8d1a706054c8a700440e2e26505bd752ab4e4efae519e5f05197.jpg",
+ "image_caption": [
+ "Fig. 10: The overview of evaluation and benchmarks in single LLM's deployment phase."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 76,
+ 502,
+ 488,
+ 652
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "text",
+ "text": "Robustness Evaluation. To systematically assess the relia-",
+ "bbox": [
+ 73,
+ 705,
+ 488,
+ 720
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "table",
+ "img_path": "images/5e09ecec36584d1cf538edd7e63fbbb1fa61ac0fa57f6f3644e5882b53355973.jpg",
+ "table_caption": [
+ "TABLE 7: Summary of LLM robustness benchmarks at the deployment stage."
+ ],
+ "table_footnote": [],
+ "table_body": "| Attacks | Method | OS | Year | Strategy | Setting | Datasets | Target Models | Metrics |
| Model Extraction | Carlini et al. [503] | Yes | 2024 | Binary Search | Black-box | None | GPTs, LLaMA, Pythia,ada, cabbage | Query&TokenCost, MSE, RMS |
| Finlayson et al. [504] | No | 2024 | Softmax Bottleneck | Black-box | None | Pythia, GPT-3.5 | Query Cost |
| Zanella et al. [505] | No | 2024 | Matrix Operations | Grey-box | SST-2, MNLI, AGNews | BERTs, XLNet | Query Cost, Acc,Agreement |
| Horwitz et al. [506] | Yes | 2024 | Spectral DeTuning | White-box | LoWRA | ViT, SD, Mistral | MSWE, SEM |
| Membership Inference | MIN-K% PROB [509] | Yes | 2023 | Probabilities | Black-box | Wikipedia | LLaMAs, Pythia, NeoX,OPT | TPR, FPR, ROC,AUC |
| MIN-K%++ [510] | Yes | 2022 | Local Maxima | Black-box | WikiMIA, MIMIR | Pythia, GPT-NeoXLLaMA, OPT, Mamba | AUROC, TPR, FPR |
| Blind [511] | Yes | 2024 | Threshold | Black-box | 8 sets | GPT-3, OpenLLaMA | AUC ROC |
| LLM-DI [512] | Yes | 2024 | Aggregation | Black-box | PILE | Pythias | AUC, p-values |
| DE-COP [513] | Yes | 2024 | Paraphrases | Black-box | arXiv:Tection, BookTection | Mistral, Mixtral, LLaMA, GPTs, Claude | AUC |
| Recall [514] | Yes | 2024 | Log-Likelihoods | Black-box | WikiMIA, MIMIR | Pythia, GPT-NeoXLLaMA, OPT, Mamba | AUC, TPR@FPR |
| Noisy [515] | No | 2024 | Embedding NGBRs | Gray-box | OpenWebText,Wikipedia | GPT-2 | TPR, FPR, AUC |
| SMIA [516] | No | 2024 | Perturbation | Gray-box | Wikipedia, FAN | Pythia, Pythia-Deduped, GPT-Neos | AUC-ROC, TPR, FPR |
| FEATAGG [517] | No | 2024 | Feature Aggregation | Black-box | ProjectGutenberg,ArXiv | OpenLLaMA | TPR@FPR, AUC |
| RAG-MIA [520] | No | 2024 | Direct Asking | Black-box | HealthCareMagic,Enron | flan, llama, mistral | TPR@FPR, AUC-ROC |
| Jailbreak | GCG [260] | Yes | 2023 | Gradient-based | White-box | Vicuna, LLaMA-2 | AdvBench | ASR, Loss |
| AmpleGCG [530] | Yes | 2024 | Hybrid-based | White-box | Vicuna, Llama-2,Mis-tral,GPTs | AdvBench | ASR, US, Diver-sity, Time |
| I-GCG [531] | Yes | 2024 | Gradient-based | White-box | AdvBench,HarmBench | VICUNA, GUANACOLLAMA, MISTRAL | ASR |
| MA-GCG [532] | Yes | 2024 | Gradient-based | White-box | AdvBench | Vicuna, Mistral | ASR, Time |
| A-GCG [533] | Yes | 2024 | Gradient-based | White-box | AdvBench | Llama2, Vicuna | ASR, Acc |
| AutoDAN-A [534] | Yes | 2023 | LLM-based | Black-box | AdvBench | Vicuna, Mistral | ASR, Recheck,PPL |
| AutoDAN-B [535] | Yes | 2023 | Gradient-based | White-box | AdvBench | Vicuna, Guanaco, Pythia | ASR, Recheck |
| PAIR [261] | Yes | 2023 | LLM-based | Black-box | JailbreakBench | Vicuna, Llama-2, GPTs,Claudes,Gemini | ASR, QPS |
| ToA [536] | Yes | 2023 | LLM-based | Black-box | AdvBench, Harm123 | Vicuna, Llama-2, PaLM-2,GPTs, Claude3, Gemini | GPT4-MetricHuman-Judge |
| PAL [537] | Yes | 2024 | LLM-based | Black-box | AdvBench | Llama-2, GPT-3.5 | ASR, Manual Labeling |
| Masterkey [538] | No | 2023 | Rephrasing | Black-box | AdvBench, Harm123 | GPTs, Bing, Bard | ASR, QSR |
| AutoDAN-Turbo [539] | Yes | 2024 | LLM-based | Black-box | Harmbench | Llama-2, Gemma, GPT-4,Gemini | ASR, StrongRE-JECT |
| FlipAttack [540] | Yes | 2025 | Rephrasing | Black-box | AdvBench, StrongRE-JECT | GPTs, Claude 3.5 Sonnet, Llama 3.1 405B, Mixtral 8x22B | ASR |
| Geneshift [541] | Yes | 2025 | LLM-based | Black-box | AdvBench | GPTs | ASR |
| Prompt Injection | IPP [542] | Yes | 2022 | Handcraft | Black-box | OpenAI Examples | text-davinci | ASR |
| Greshake et al. [543] | Yes | 2023 | Data Poisoning | Black-box | None | text-davinci, GPT-4 | None |
| HOUYI [544] | Yes | 2023 | Components Asmbl | Black-box | Five Queries | SUPERTOOLS | Manual |
| Yan et al. [130] | Yes | 2023 | Poisoning | Black-box | Several Cases | Alpaca | Ngt, Pst, Ocrc |
| TT [545] | No | 2023 | Game | Black-box | Tensor Trust | GPTs, Claudes, PaLM, LLaMAs | Robustness Rate |
| JudgeDeceiver [546] | Yes | 2024 | Gradient-based | White-box | MT-Bench, LLMBarMRPC, Jfleg, HSOL,RTE, SST2, SMS | Mistral, Openchat, Llamas | ACC, ASR, PACKEY-E, LM-E |
| AUPI [547] | Yes | 2024 | Gradient-based | White-box | MRPC, Jfleg, HSOL,RTE, SST2, SMS | Llama2 | ASR |
| AUTOHIJACKER [548] | No | 2024 | LLM-based | Black-box | AgentDojo, OPI | Llama, Command-R,GPTs | ASR |
| Data Extraction | zlib [108] | Yes | 2020 | Generate & Inference | Black-box | Top-n, Temperature, Internet | GPT-2 | 6 metrics |
| AutoSklearn [549] | No | 2023 | Greedy, Contrastive, Beam decoding | Black-box | Pile | GPT-Neo | Precision, Recall,R@FPR |
| DECOM [550] | No | 2024 | Decomposition | Black-box | NYT, WSJ | Frontiers | TRM, EMP,BITAP |
| Context [551] | No | 2022 | Context, Zero-shot,Few-shot | Black-box | Enron Corpus | GPT-Neo | Acc |
| ETHICIST [552] | Yes | 2023 | Prompt Tuning | Gray-box | LM-Extraction | GPT-Neo | Recall |
| Pli-compass [553] | No | 2024 | Grounding | Black-box | Enron email | GPT-J | Extraction Rate |
| DSP [554] | No | 2024 | Dynamic Soft Prompting | Black-box | LMEB, The Stack | GPT-Neo, Pythia, Star-CoderBase | EER, FER, PPL |
| PWB [555] | Yes | 2024 | Gradient-based | White-box | Pile | Pythia, Llama | Precision, AUC,TPR |
| Prompt Stealing | Sha et al. [556] | No | 2024 | LLM-based | Black-box | RetrievalQA,AlpacaGPT4 | ChatGPT, LLaMA | Acc, Precision, Recall, AUC |
| output2prompt [557] | Yes | 2024 | LLM-based | Black-box | 3 User & 3 SystemPrompts | Llamas, GPTs | BLEU, CS, Preci-sion, Recall |
| PRSA [558] | No | 2024 | Output Difference | Black-box | Category18 | GPTs | BLEU, FastKAS-SIM, JS |
"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.075,
+ 0.032,
+ 0.422,
+ 0.045
+ ],
+ "angle": 0,
+ "content": "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.907,
+ 0.034,
+ 0.921,
+ 0.043
+ ],
+ "angle": 0,
+ "content": "21"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.086,
+ 0.053,
+ 0.491,
+ 0.112
+ ],
+ "angle": 0,
+ "content": "ECLIPSE [588] uses an LLM as a suffix generator and optimizer. PAL [537] proposes an online proxy model (which is used for adversarial prompt generation) training pipeline."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.068,
+ 0.112,
+ 0.492,
+ 0.215
+ ],
+ "angle": 0,
+ "content": "\\(\\Rightarrow\\) Others. EnJa [589] proposes to ensemble prompt and token-level attack methods via a template-based connector. AmpleGCG [530] first collects lots of successful suffixes and then trains the generative model to generate a specific suffix for a given unsafe prompt. Zhao et al. [590] targets the scenario where the decoding process of target LLM is assisted with smaller models' guidance."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.068,
+ 0.221,
+ 0.492,
+ 0.324
+ ],
+ "angle": 0,
+ "content": "Prompt Injection Attacks. Prompt injection is a vulnerability where an attacker manipulates the input prompts of LLMs to force them to generate a specific output, which is usually out of the range for normal use (e.g., goal hijacking and prompt leaking [542]), often by injecting malicious text or commands into the input field. Attackers can employ a variety of techniques to carry out such attacks."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.068,
+ 0.33,
+ 0.493,
+ 0.593
+ ],
+ "angle": 0,
+ "content": "\\(\\Rightarrow\\) Direct Prompt Injection. Perez et al. [542] directly inject handcrafted adversarial prompts into inputs to misalign the language model. HOUYI [544] proposes an injection generation framework which includes three components. Yan et al. [130] utilize LLMs to generate diverse trigger instructions that implicitly capture the characteristics of trigger scenarios. TENSOR TRUST leverages the TENSOR TRUST web game to generate a large-scale dataset and benchmark [545]. AUPI [547] adopts a gradient-based optimization method, specifically, a momentum-enhanced optimization algorithm, to generate universal prompt injection data. Upadhayay et al. [591] argue that LLMs suffer from cognitive overload and propose to use in-context learning to jailbreak LLMs through deliberately designed prompts that induce cognitive overload. Kwon et al. [592] circumvent security policies by substituting sensitive words—likely to be rejected by the language model—with mathematical functions."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.068,
+ 0.593,
+ 0.493,
+ 0.884
+ ],
+ "angle": 0,
+ "content": "\\(\\nLeftrightarrow\\) Indirect Prompt Injection. Greshake et al. [543] propose to indirectly inject prompts into the data that are likely to be retrieved. Bagdasaryan et al. [593] design a prompt injection attack against multi-modal LLMs, by generating an adversarial perturbation corresponding to the prompt and blending it into an image or audio recording. Neural Exec [594] designs a multi-stage preprocessing pipeline for cases like Retrieval-Augmented Generation (RAG)-based applications. PoisonedAlign [595] boosts the success of prompt injection attacks by strategically creating poisoned alignment samples in the LLM's alignment process. TPIA [596] crafts non-functional perturbations that contain malicious information and inserts them into the victim's code context by spreading them into potentially used dependencies like packages or RAG's knowledge base. F2A [597] proposes to use feign security detection agents to bypass the defense mechanism of LLMs. AUTOHIJACKER [548] uses a batch-based optimization framework to handle sparse feedback and leverages a trainable memory to enable effective generation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.068,
+ 0.884,
+ 0.492,
+ 0.944
+ ],
+ "angle": 0,
+ "content": "Different Settings. JudgeDeceiver uses gradient-based optimization to inject LLM-as-a-Judge scenarios [546]. Pedro et al. [598] study the risk of injections targeting web applications based on the Langchain framework. Lee et"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.068,
+ 0.33,
+ 0.493,
+ 0.944
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.516,
+ 0.053,
+ 0.923,
+ 0.157
+ ],
+ "angle": 0,
+ "content": "al. [599] propose a human-AI collaborative framework to explore the potential of prompt injection against federated military LLMs. PROMPT INFECTION [600] proposes to make malicious prompts self-replicate across interconnected agents in multi-agent systems. Zhang et al. [601] explore the risk of prompt injection in LLM-integrated systems like LLM-integrated mobile robotic systems."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.504,
+ 0.163,
+ 0.924,
+ 0.235
+ ],
+ "angle": 0,
+ "content": "Data Extraction Attacks. Data extraction attacks try to figure out the personally identifiable information (PII) that is used to train the LLMs [108]. It starts from sufficient-length prefixes to perform extraction and additional measures to determine if extracted texts are valid."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.503,
+ 0.243,
+ 0.924,
+ 0.606
+ ],
+ "angle": 0,
+ "content": "\\(\\nLeftrightarrow\\) Methods. In the beginning work [108], the proposed extraction process contains two stages \"generate-then-rank\": sampling potentially memorized examples and membership inference. It proposes a temperature-decaying method to sample more diverse examples and use surrogate models to infer the membership. After that, Al-Kaswan et al. [549] propose using greedy, contrastive, and beam decoding strategies to generate examples and use a classifier to infer the membership. Su et al. [550] propose an instruction decomposition technique to extract fragments of training data gradually. Huang et al. [551] extensively explore the effect of context, zero-shot, and few-shot methods in extracting the personal email address. ETHICIST proposes a smoothing loss and a calibrated confidence estimation method to extract the suffix and measure the confidence [552]. Nakka et al. [553] improves the extraction performance by grounding the prefix of the manually constructed extraction prompt with in-domain data. Wang et al. [554] propose to train a transformer-based generator to produce dynamic, prefix-dependent soft prompts. Ozdayi et al. [105] introduce an approach that uses prompt tuning to control the extraction rates of memorized content. Meng et al. [602] propose a two-stage method, i.e., collection and ranking, to recover PPI when PII entities have been masked."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.502,
+ 0.608,
+ 0.924,
+ 0.927
+ ],
+ "angle": 0,
+ "content": "Different Settings. Some works also explore the risk of data leakage in novel settings. Wang et al. [555] study the probability of data extraction in fine-tuning settings and Bargav et al. [603, 604] extract the training data by comparing the output difference before and after the fine-tuning. Jiang et al. [605, 606, 607] propose to extract the private Retrieval-Augmented Generation (RAG) documents. Peng et al. [608] extract the private RAG documents by poisoning in the fine-tuning process. Nasr et al. [107] explore the potential risk of data extraction for the aligned production language models. Panda et al. [609] extract the fine-tuning secret data by poisoning the pertaining dataset. Lu et al. [610] propose to extract PII from an aligned model with model merging. Chen et al. [611] find that fine-tuning can recover the forgotten PIIs in pretraining data. Panchendrarajan et al. [612] propose to extract the whole private training data in the fine-tuning process. Rashid et al. [613] propose selective weight tampering to explore PPI leakage in Federated Language Models. Dentan et al. [614] extract data from layout-aware document understanding models like unimodal or bimodal models."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.501,
+ 0.928,
+ 0.923,
+ 0.944
+ ],
+ "angle": 0,
+ "content": "Different Applications. Leveraging the abnormally high"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.501,
+ 0.243,
+ 0.924,
+ 0.944
+ ],
+ "angle": 0,
+ "content": null
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.075,
+ 0.033,
+ 0.422,
+ 0.045
+ ],
+ "angle": 0,
+ "content": "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.907,
+ 0.034,
+ 0.922,
+ 0.043
+ ],
+ "angle": 0,
+ "content": "22"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.086,
+ 0.054,
+ 0.493,
+ 0.287
+ ],
+ "angle": 0,
+ "content": "token probabilities, some works utilize the memorization of LLMs to extract the fingerprint or steganography [615]. Al-Kaswan et al. [616] explore memorization in large language models for code and find that code models memorize training data at a lower rate than natural language models. Nie et al. [617] utilize the token-level features derived from the identified characteristics to decode the PII. Lehman et al. [618] reveal the risk of Electronic Health Records leakage of LLMs. Diera et al. [619] conduct experiments to assess the PII leakage of fine-tuned BERT models and found that Differential Privacy (DP) has a negative effect when deployed in fine-tuning. Zhang et al. [620] propose data extraction attacks against text classification with transformers. Huang et al. [621] propose an evaluation tool, i.e. HCR, to assess the PPI leakage in Neural Code Completion Tools."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.069,
+ 0.288,
+ 0.493,
+ 0.476
+ ],
+ "angle": 0,
+ "content": "\\(\\nrightarrow\\) Factor Assessment. Some work studies the factors of data extraction including decoding schemes, model sizes, prefix lengths, partial sequence leakages, and token positions [622, 623]. Yash et al. [624] explore the effects of prompt sensitivity and access to multiple checkpoints to extraction attacks. Staab et al. [625] construct a dataset consisting of real Reddit profiles to extract personal attributes. Xu et al. [626] conduct experiments to evaluate the factors of different suffix generation methods and different membership inference attacks in extraction performance. Karamolegkou et al. [627] evaluate the effect of model structure, data type, probing strategies, and metrics."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.072,
+ 0.478,
+ 0.493,
+ 0.741
+ ],
+ "angle": 0,
+ "content": "Prompt Stealing Attacks. Given that crafting effective prompts requires significant engineering effort and can be considered valuable intellectual property (IP), promptstealing attacks aim to compromise this IP by reconstructing prompts from generated responses [556, 557, 558]. These generation effects are often used to attract prospective prospective buyers. Sha et al. [556] pioneer this approach by collecting a dataset and training classifiers to predict prompt parameters—such as whether the prompt is direct, role-based, or in-context. They then used a large language model (LLM) to reconstruct the prompt. Similarly, Zhang et al. [557] trained an LLM on output-prompt pairs to directly infer the original prompt, while Yang et al. [558] leveraged generation differences to refine surrogate prompts. However, recovering the original prompt solely from the output is challenging. Out of this, Zheng et al. [628] propose a timing-based side-channel method to infer the prompt during inference."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.074,
+ 0.751,
+ 0.395,
+ 0.766
+ ],
+ "angle": 0,
+ "content": "6.1.2 Defensive Mechanisms in Deployment"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.073,
+ 0.768,
+ 0.491,
+ 0.914
+ ],
+ "angle": 0,
+ "content": "In Subsubsection 6.1.1, we analyzed various attack scenarios targeting individual LLM deployments. However, in real-world applications, defense mechanisms are not designed as isolated, one-to-one countermeasures against specific attacks. Instead, they follow fundamental security principles to establish a systematic defense framework, as illustrated in Figure 9. This framework integrates multiple layers of protection, ensuring resilience against a wide range of adversarial threats while maintaining model usability and efficiency."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.073,
+ 0.914,
+ 0.492,
+ 0.945
+ ],
+ "angle": 0,
+ "content": "Input Preprocessing Defenses Input preprocessing serves as the first line of defense in LLM deployment, aiming to"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.508,
+ 0.059,
+ 0.764,
+ 0.072
+ ],
+ "angle": 0,
+ "content": "Defensive Mechanisms in Deployment"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.508,
+ 0.072,
+ 0.916,
+ 0.16
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.505,
+ 0.169,
+ 0.922,
+ 0.2
+ ],
+ "angle": 0,
+ "content": "Fig. 9: The overview of attacks in single LLM's deployment phase."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.504,
+ 0.226,
+ 0.922,
+ 0.254
+ ],
+ "angle": 0,
+ "content": "detect and neutralize adversarial inputs before they reach the model."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.503,
+ 0.256,
+ 0.923,
+ 0.504
+ ],
+ "angle": 0,
+ "content": "Attack Detection & Identification: Effective input filtering [629, 630] begins with attack detection [631], which identifies adversarial prompts through statistical [632], structural [633], or behavioral inconsistencies [634]. Gradient-based detection methods [635] leverage safety-critical gradient analysis and loss landscape exploration to uncover jailbreak prompts that manipulate LLM behavior. These approaches identify adversarial inputs [636, 637] by analyzing how small perturbations [638] affect model outputs, detecting highly sensitive or misaligned gradients that indicate targeted attacks. Perplexity-based methods [632, 632] measure the probability distribution of input sequences, flagging atypical or low-likelihood prompts as potential adversarial inputs. These techniques are particularly effective in detecting prompt injection and adversarial perturbations, where crafted prompts deviate significantly from natural language distributions."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.503,
+ 0.505,
+ 0.923,
+ 0.621
+ ],
+ "angle": 0,
+ "content": "Beyond individual heuristics, universal detection frameworks [639] integrate multiple detection strategies to counter diverse attack vectors, including prompt injection [640], backdoor manipulations [641], and adversarial attacks [637]. These frameworks employ ensemble-based filtering mechanisms, combining gradient analysis [642], perplexity estimation [643], and syntactic evaluation for generalized attack resilience."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.503,
+ 0.622,
+ 0.923,
+ 0.87
+ ],
+ "angle": 0,
+ "content": "Semantic & Behavioral Analysis: Attack detection alone is insufficient, as certain adversarial inputs may bypass traditional filtering mechanisms. Semantic [644] and behavioral analysis enhance input preprocessing by evaluating linguistic intent and model alignment. Self-examination techniques allow LLMs [645, 646] to assess whether they are being manipulated, leveraging auxiliary reasoning steps to detect deceptive prompts. Alignment-based verification [647] ensures that the model's responses remain consistent with its safety objectives [330], identifying inputs that subtly nudge the model toward policy violations or ethical misalignment. Intention analysis [648, 649] further refines input filtering by discerning subtle manipulations designed to bypass explicit security checks. Unlike token-level detection, which flags overtly adversarial inputs, intention-aware defenses analyze the semantic structure and purpose of the input to preemptively reject jailbreak attempts."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.504,
+ 0.87,
+ 0.922,
+ 0.945
+ ],
+ "angle": 0,
+ "content": "Adversarial Defense & Mitigation: When detection and behavioral analysis fail to fully neutralize adversarial inputs, robustness-enhancing techniques [647] mitigate their effects by reducing model susceptibility to manipulation [334, 650]. Semantic smoothing [651, 652] techniques"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.075,
+ 0.032,
+ 0.422,
+ 0.045
+ ],
+ "angle": 0,
+ "content": "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.907,
+ 0.034,
+ 0.923,
+ 0.044
+ ],
+ "angle": 0,
+ "content": "23"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.073,
+ 0.054,
+ 0.493,
+ 0.142
+ ],
+ "angle": 0,
+ "content": "introduce controlled randomness into LLM responses, reducing the model's sensitivity to adversarial perturbations and preventing reliable jailbreak execution. By stabilizing decision boundaries [653], these methods enhance resistance against prompt manipulation strategies that exploit response predictability."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.073,
+ 0.142,
+ 0.493,
+ 0.258
+ ],
+ "angle": 0,
+ "content": "Preemptive input transformations [654], such as back-translation [655] or paraphrasing, modify incoming queries [651] while preserving semantic intent, disrupting adversarial structures embedded within malicious prompts. Data augmentation [656] and adversarial training further strengthen model robustness by exposing LLMs to adversarial prompts during training, forcing them to learn invariances that reduce their vulnerability to real-world attacks."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.073,
+ 0.258,
+ 0.493,
+ 0.389
+ ],
+ "angle": 0,
+ "content": "Output Filtering Mechanisms. Output filtering mechanisms [212, 657] serve as a critical safeguard in LLM deployment, ensuring that generated responses comply with safety constraints while preserving informativeness. Unlike input preprocessing, which aims to prevent adversarial prompts from reaching the model, output filtering mitigates harmful content post-generation. Existing approaches primarily follow three paradigms: rule-based constraints, generative adversarial filtering, and toxicity detection."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.39,
+ 0.492,
+ 0.549
+ ],
+ "angle": 0,
+ "content": "Rule-based mechanisms [658] impose predefined constraints on model outputs, preventing the generation of harmful, unethical, or undesired content. Programmable guardrails [659] offer a structured framework where developers can enforce response filtering, topic restriction, and ethical alignment. These methods often integrate reinforcement learning from human feedback [155] or rule-based reward [660] modeling to refine output safety. While effective at handling explicit violations, static rule-based methods struggle with nuanced adversarial prompts and subtle misalignments."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.073,
+ 0.55,
+ 0.493,
+ 0.738
+ ],
+ "angle": 0,
+ "content": "To address these limitations, generative adversarial filtering [661] leverages self-critique [662, 663], ensemble detection, and dynamic response evaluation [664]. Self-rectification mechanisms [663, 665] enable LLMs to critique their own outputs and refine responses through iterative refinement. Additionally, ensemble-based [666] moderation models aggregate predictions from multiple LLMs, improving robustness against circumvention techniques. Adaptive filtering frameworks [667] employ perplexity-based assessments and adversarial perturbation detection to flag responses deviating from expected linguistic patterns, enhancing their resilience against jailbreak attempts [668, 669] and toxic content injection."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.073,
+ 0.739,
+ 0.493,
+ 0.913
+ ],
+ "angle": 0,
+ "content": "Toxicity detection [670, 671, 672] and content moderation [673, 674, 675, 676] further reinforce output safety by identifying and mitigating hate speech [677], misinformation, and other harmful content. Supervised finetuning adapts LLMs to recognize undesirable patterns, while classifier-based detection models [678] filter responses in real-time. Some approaches introduce debiasing strategies, such as controlled decoding [679, 680] and anti-expert guidance [681], to suppress toxic outputs without sacrificing response diversity. However, these methods face challenges in balancing false positives and false negatives, particularly in ambiguous or context-dependent cases."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.074,
+ 0.914,
+ 0.493,
+ 0.943
+ ],
+ "angle": 0,
+ "content": "The effectiveness of output filtering hinges on its ability to balance strict control with linguistic flexibility, ensur"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.503,
+ 0.054,
+ 0.923,
+ 0.213
+ ],
+ "angle": 0,
+ "content": "ing that models remain both safe and practically useful. A hybrid approach combining rule-based safeguards, self-correcting mechanisms, and adaptive toxicity moderation is essential to achieving robust and scalable LLM deployment. Robust Prompt Engineering. Robust prompt engineering aims to enhance LLM safety by designing input prompts that resist adversarial manipulation [682], protect sensitive data, and mitigate harmful outputs—all [683] without modifying model parameters. These strategies act at the interaction level, offering lightweight and model-agnostic protection."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.504,
+ 0.214,
+ 0.924,
+ 0.461
+ ],
+ "angle": 0,
+ "content": "Recent efforts have introduced prompt optimization techniques grounded in adversarial robustness, including embedding-space manipulation and defensive objective alignment. Methods such as Robust Prompt Optimization [684] and Prompt Adversarial Tuning generate transferable suffixes [668] or prefix [685] embeddings to guide model behavior [686] under attack [687], effectively lowering jailbreak success rates while preserving task performance. Similarly, goal prioritization frameworks [688] enforce inference-time objective consistency, dynamically resolving conflict between user instructions and safety constraints without requiring access to malicious samples. Complementary to these strategies, patch-based methods integrate interpretable suffixes or structured self-reminders [689] into prompts, reducing the model's susceptibility to coercive inputs through lightweight, modular defenses."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.504,
+ 0.462,
+ 0.924,
+ 0.55
+ ],
+ "angle": 0,
+ "content": "Structural manipulation approaches [690] neutralize adversarial intent through prompt rewriting. Spotlighting [691] injects source-attribute signals to counter indirect prompt injection, while inverse prompt engineering [692] repurposes attack data to generate task-specific defensive prompts under the principle of least privilege."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.504,
+ 0.55,
+ 0.924,
+ 0.724
+ ],
+ "angle": 0,
+ "content": "Privacy-preserving prompt [693] design introduces formal guarantees through differential privacy. Approaches like DP-Prompt [694] and stochastic gradient masking [695] reduce information leakage from prompts without harming performance. Desensitization and directional control of incontext representations offer additional privacy protections during prompt construction. Prompt engineering [579, 696] also helps mitigate societal risks. Chain-of-thought prompting and guided templates reduce gender bias [697] in reasoning tasks, while prompt learning [698] improves toxicity detection and generation control [699, 700], often surpassing specialized models in efficiency and generalization."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.504,
+ 0.724,
+ 0.923,
+ 0.812
+ ],
+ "angle": 0,
+ "content": "Finally, systematic prompt optimization methods [701, 702] aim to generalize prompt robustness across tasks and domains. Techniques like BATPrompt [703] and StraGo [704] use adversarial simulation and strategic decomposition to refine prompts iteratively, improving both resilience and effectiveness under variable inputs."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.504,
+ 0.812,
+ 0.924,
+ 0.943
+ ],
+ "angle": 0,
+ "content": "System-level Security Controls. System-level defenses [705] enhance LLM deployment by optimizing inference, enforcing alignment, isolating untrusted inputs, and securing the supply chain. Systems like Petals [706], Sarathi-Serve [707], and DistServe [708] restructure computation to improve throughput and latency, while TriForce [709], Medusa [710] MagicDec [711] accelerate generation via speculative decoding and structural compression. Parallel frameworks such as DeepSpeed-FastGen [712] and SpecExec [713] further boost"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.075,
+ 0.032,
+ 0.422,
+ 0.045
+ ],
+ "angle": 0,
+ "content": "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.907,
+ 0.034,
+ 0.922,
+ 0.043
+ ],
+ "angle": 0,
+ "content": "24"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.073,
+ 0.054,
+ 0.314,
+ 0.068
+ ],
+ "angle": 0,
+ "content": "efficiency with minimal overhead."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.072,
+ 0.068,
+ 0.491,
+ 0.171
+ ],
+ "angle": 0,
+ "content": "Runtime alignment methods [714] adapt model behavior through cross-model guidance or token-level reward modeling. Systems such as SelfDefend [715] and Gradient Cuff [716] detect unsafe generation by monitoring agreement across models or loss landscapes, while Spotlighting [691] inserts provenance signals to mitigate indirect prompt injection."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.072,
+ 0.171,
+ 0.492,
+ 0.245
+ ],
+ "angle": 0,
+ "content": "Access isolation is achieved through policy enforcement [717] and system wrappers [688]. At the supply level, tools like MalHug [718] identify poisoned models, while system audits reveal sandbox and plugin vulnerabilities, highlighting the need for end-to-end secure deployment."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.072,
+ 0.245,
+ 0.491,
+ 0.378
+ ],
+ "angle": 0,
+ "content": "LLM-based guard models utilize lightweight LLMs like Llama Guard [330], Aegis Guard [719, 720], WildGuard [721], and ShieldGemma [722] to moderate both the input and output of the victim LLMs. However, they are purely classifiers. To solve this problem, the first reasoning-based guard model named GuardReasoner [723] is proposed to improve the performance, explainability, and generalization ability via learning to reason. It brings new opportunities for the safety of large-scale reasoning models [724]."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.074,
+ 0.393,
+ 0.428,
+ 0.408
+ ],
+ "angle": 0,
+ "content": "6.1.3 Evaluation and Benchmarks in Deployment"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.072,
+ 0.412,
+ 0.492,
+ 0.487
+ ],
+ "angle": 0,
+ "content": "To assess the reliability and safety of LLMs after deployment, evaluation efforts focus on several key dimensions and risk types, as illustrated in Figure 10. These dimensions guide the design of systematic benchmarks and metrics tailored for real-world deployment settings."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.077,
+ 0.503,
+ 0.49,
+ 0.653
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.089,
+ 0.662,
+ 0.478,
+ 0.693
+ ],
+ "angle": 0,
+ "content": "Fig. 10: The overview of evaluation and benchmarks in single LLM's deployment phase."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.074,
+ 0.706,
+ 0.49,
+ 0.722
+ ],
+ "angle": 0,
+ "content": "Robustness Evaluation. To systematically assess the relia-"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.079,
+ 0.733,
+ 0.486,
+ 0.763
+ ],
+ "angle": 0,
+ "content": "TABLE 7: Summary of LLM robustness benchmarks at the deployment stage."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.076,
+ 0.767,
+ 0.49,
+ 0.94
+ ],
+ "angle": 0,
+ "content": "| Attacks | Method | OS | Year | Strategy | Setting | Datasets | Target Models | Metrics |
| Model Extraction | Carlini et al. [503] | Yes | 2024 | Binary Search | Black-box | None | GPTs, LLaMA, Pythia,ada, cabbage | Query&TokenCost, MSE, RMS |
| Finlayson et al. [504] | No | 2024 | Softmax Bottleneck | Black-box | None | Pythia, GPT-3.5 | Query Cost |
| Zanella et al. [505] | No | 2024 | Matrix Operations | Grey-box | SST-2, MNLI, AGNews | BERTs, XLNet | Query Cost, Acc,Agreement |
| Horwitz et al. [506] | Yes | 2024 | Spectral DeTuning | White-box | LoWRA | ViT, SD, Mistral | MSWE, SEM |
| Membership Inference | MIN-K% PROB [509] | Yes | 2023 | Probabilities | Black-box | Wikipedia | LLaMAs, Pythia, NeoX,OPT | TPR, FPR, ROC,AUC |
| MIN-K%++ [510] | Yes | 2022 | Local Maxima | Black-box | WikiMIA, MIMIR | Pythia, GPT-NeoXLLaMA, OPT, Mamba | AUROC, TPR, FPR |
| Blind [511] | Yes | 2024 | Threshold | Black-box | 8 sets | GPT-3, OpenLLaMA | AUC ROC |
| LLM-DI [512] | Yes | 2024 | Aggregation | Black-box | PILE | Pythias | AUC, p-values |
| DE-COP [513] | Yes | 2024 | Paraphrases | Black-box | arXiv:Tection, BookTection | Mistral, Mixtral, LLaMA, GPTs, Claude | AUC |
| Recall [514] | Yes | 2024 | Log-Likelihoods | Black-box | WikiMIA, MIMIR | Pythia, GPT-NeoXLLaMA, OPT, Mamba | AUC, TPR@FPR |
| Noisy [515] | No | 2024 | Embedding NGBRs | Gray-box | OpenWebText,Wikipedia | GPT-2 | TPR, FPR, AUC |
| SMIA [516] | No | 2024 | Perturbation | Gray-box | Wikipedia, FAN | Pythia, Pythia-Deduped, GPT-Neos | AUC-ROC, TPR, FPR |
| FEATAGG [517] | No | 2024 | Feature Aggregation | Black-box | ProjectGutenberg,ArXiv | OpenLLaMA | TPR@FPR, AUC |
| RAG-MIA [520] | No | 2024 | Direct Asking | Black-box | HealthCareMagic,Enron | flan, llama, mistral | TPR@FPR, AUC-ROC |
| Jailbreak | GCG [260] | Yes | 2023 | Gradient-based | White-box | Vicuna, LLaMA-2 | AdvBench | ASR, Loss |
| AmpleGCG [530] | Yes | 2024 | Hybrid-based | White-box | Vicuna, Llama-2,Mis-tral,GPTs | AdvBench | ASR, US, Diver-sity, Time |
| I-GCG [531] | Yes | 2024 | Gradient-based | White-box | AdvBench,HarmBench | VICUNA, GUANACOLLAMA, MISTRAL | ASR |
| MA-GCG [532] | Yes | 2024 | Gradient-based | White-box | AdvBench | Vicuna, Mistral | ASR, Time |
| A-GCG [533] | Yes | 2024 | Gradient-based | White-box | AdvBench | Llama2, Vicuna | ASR, Acc |
| AutoDAN-A [534] | Yes | 2023 | LLM-based | Black-box | AdvBench | Vicuna, Mistral | ASR, Recheck,PPL |
| AutoDAN-B [535] | Yes | 2023 | Gradient-based | White-box | AdvBench | Vicuna, Guanaco, Pythia | ASR, Recheck |
| PAIR [261] | Yes | 2023 | LLM-based | Black-box | JailbreakBench | Vicuna, Llama-2, GPTs,Claudes,Gemini | ASR, QPS |
| ToA [536] | Yes | 2023 | LLM-based | Black-box | AdvBench, Harm123 | Vicuna, Llama-2, PaLM-2,GPTs, Claude3, Gemini | GPT4-MetricHuman-Judge |
| PAL [537] | Yes | 2024 | LLM-based | Black-box | AdvBench | Llama-2, GPT-3.5 | ASR, Manual Labeling |
| Masterkey [538] | No | 2023 | Rephrasing | Black-box | AdvBench, Harm123 | GPTs, Bing, Bard | ASR, QSR |
| AutoDAN-Turbo [539] | Yes | 2024 | LLM-based | Black-box | Harmbench | Llama-2, Gemma, GPT-4,Gemini | ASR, StrongRE-JECT |
| FlipAttack [540] | Yes | 2025 | Rephrasing | Black-box | AdvBench, StrongRE-JECT | GPTs, Claude 3.5 Sonnet, Llama 3.1 405B, Mixtral 8x22B | ASR |
| Geneshift [541] | Yes | 2025 | LLM-based | Black-box | AdvBench | GPTs | ASR |
| Prompt Injection | IPP [542] | Yes | 2022 | Handcraft | Black-box | OpenAI Examples | text-davinci | ASR |
| Greshake et al. [543] | Yes | 2023 | Data Poisoning | Black-box | None | text-davinci, GPT-4 | None |
| HOUYI [544] | Yes | 2023 | Components Asmbl | Black-box | Five Queries | SUPERTOOLS | Manual |
| Yan et al. [130] | Yes | 2023 | Poisoning | Black-box | Several Cases | Alpaca | Ngt, Pst, Ocrc |
| TT [545] | No | 2023 | Game | Black-box | Tensor Trust | GPTs, Claudes, PaLM, LLaMAs | Robustness Rate |
| JudgeDeceiver [546] | Yes | 2024 | Gradient-based | White-box | MT-Bench, LLMBarMRPC, Jfleg, HSOL,RTE, SST2, SMS | Mistral, Openchat, Llamas | ACC, ASR, PACKEY-E, LM-E |
| AUPI [547] | Yes | 2024 | Gradient-based | White-box | MRPC, Jfleg, HSOL,RTE, SST2, SMS | Llama2 | ASR |
| AUTOHIJACKER [548] | No | 2024 | LLM-based | Black-box | AgentDojo, OPI | Llama, Command-R,GPTs | ASR |
| Data Extraction | zlib [108] | Yes | 2020 | Generate & Inference | Black-box | Top-n, Temperature, Internet | GPT-2 | 6 metrics |
| AutoSklearn [549] | No | 2023 | Greedy, Contrastive, Beam decoding | Black-box | Pile | GPT-Neo | Precision, Recall,R@FPR |
| DECOM [550] | No | 2024 | Decomposition | Black-box | NYT, WSJ | Frontiers | TRM, EMP,BITAP |
| Context [551] | No | 2022 | Context, Zero-shot,Few-shot | Black-box | Enron Corpus | GPT-Neo | Acc |
| ETHICIST [552] | Yes | 2023 | Prompt Tuning | Gray-box | LM-Extraction | GPT-Neo | Recall |
| Pli-compass [553] | No | 2024 | Grounding | Black-box | Enron email | GPT-J | Extraction Rate |
| DSP [554] | No | 2024 | Dynamic Soft Prompting | Black-box | LMEB, The Stack | GPT-Neo, Pythia, Star-CoderBase | EER, FER, PPL |
| PWB [555] | Yes | 2024 | Gradient-based | White-box | Pile | Pythia, Llama | Precision, AUC,TPR |
| Prompt Stealing | Sha et al. [556] | No | 2024 | LLM-based | Black-box | RetrievalQA,AlpacaGPT4 | ChatGPT, LLaMA | Acc, Precision, Recall, AUC |
| output2prompt [557] | Yes | 2024 | LLM-based | Black-box | 3 User & 3 SystemPrompts | Llamas, GPTs | BLEU, CS, Preci-sion, Recall |
| PRSA [558] | No | 2024 | Output Difference | Black-box | Category18 | GPTs | BLEU, FastKAS-SIM, JS |
+
+ECLIPSE [588] uses an LLM as a suffix generator and optimizer. PAL [537] proposes an online proxy model (which is used for adversarial prompt generation) training pipeline.
+
+$\Rightarrow$ Others. EnJa [589] proposes to ensemble prompt and token-level attack methods via a template-based connector. AmpleGCG [530] first collects lots of successful suffixes and then trains the generative model to generate a specific suffix for a given unsafe prompt. Zhao et al. [590] targets the scenario where the decoding process of target LLM is assisted with smaller models' guidance.
+
+Prompt Injection Attacks. Prompt injection is a vulnerability where an attacker manipulates the input prompts of LLMs to force them to generate a specific output, which is usually out of the range for normal use (e.g., goal hijacking and prompt leaking [542]), often by injecting malicious text or commands into the input field. Attackers can employ a variety of techniques to carry out such attacks.
+
+$\Rightarrow$ Direct Prompt Injection. Perez et al. [542] directly inject handcrafted adversarial prompts into inputs to misalign the language model. HOUYI [544] proposes an injection generation framework which includes three components. Yan et al. [130] utilize LLMs to generate diverse trigger instructions that implicitly capture the characteristics of trigger scenarios. TENSOR TRUST leverages the TENSOR TRUST web game to generate a large-scale dataset and benchmark [545]. AUPI [547] adopts a gradient-based optimization method, specifically, a momentum-enhanced optimization algorithm, to generate universal prompt injection data. Upadhayay et al. [591] argue that LLMs suffer from cognitive overload and propose to use in-context learning to jailbreak LLMs through deliberately designed prompts that induce cognitive overload. Kwon et al. [592] circumvent security policies by substituting sensitive words—likely to be rejected by the language model—with mathematical functions.
+$\nLeftrightarrow$ Indirect Prompt Injection. Greshake et al. [543] propose to indirectly inject prompts into the data that are likely to be retrieved. Bagdasaryan et al. [593] design a prompt injection attack against multi-modal LLMs, by generating an adversarial perturbation corresponding to the prompt and blending it into an image or audio recording. Neural Exec [594] designs a multi-stage preprocessing pipeline for cases like Retrieval-Augmented Generation (RAG)-based applications. PoisonedAlign [595] boosts the success of prompt injection attacks by strategically creating poisoned alignment samples in the LLM's alignment process. TPIA [596] crafts non-functional perturbations that contain malicious information and inserts them into the victim's code context by spreading them into potentially used dependencies like packages or RAG's knowledge base. F2A [597] proposes to use feign security detection agents to bypass the defense mechanism of LLMs. AUTOHIJACKER [548] uses a batch-based optimization framework to handle sparse feedback and leverages a trainable memory to enable effective generation.
+Different Settings. JudgeDeceiver uses gradient-based optimization to inject LLM-as-a-Judge scenarios [546]. Pedro et al. [598] study the risk of injections targeting web applications based on the Langchain framework. Lee et
+
+al. [599] propose a human-AI collaborative framework to explore the potential of prompt injection against federated military LLMs. PROMPT INFECTION [600] proposes to make malicious prompts self-replicate across interconnected agents in multi-agent systems. Zhang et al. [601] explore the risk of prompt injection in LLM-integrated systems like LLM-integrated mobile robotic systems.
+
+Data Extraction Attacks. Data extraction attacks try to figure out the personally identifiable information (PII) that is used to train the LLMs [108]. It starts from sufficient-length prefixes to perform extraction and additional measures to determine if extracted texts are valid.
+
+$\nLeftrightarrow$ Methods. In the beginning work [108], the proposed extraction process contains two stages "generate-then-rank": sampling potentially memorized examples and membership inference. It proposes a temperature-decaying method to sample more diverse examples and use surrogate models to infer the membership. After that, Al-Kaswan et al. [549] propose using greedy, contrastive, and beam decoding strategies to generate examples and use a classifier to infer the membership. Su et al. [550] propose an instruction decomposition technique to extract fragments of training data gradually. Huang et al. [551] extensively explore the effect of context, zero-shot, and few-shot methods in extracting the personal email address. ETHICIST proposes a smoothing loss and a calibrated confidence estimation method to extract the suffix and measure the confidence [552]. Nakka et al. [553] improves the extraction performance by grounding the prefix of the manually constructed extraction prompt with in-domain data. Wang et al. [554] propose to train a transformer-based generator to produce dynamic, prefix-dependent soft prompts. Ozdayi et al. [105] introduce an approach that uses prompt tuning to control the extraction rates of memorized content. Meng et al. [602] propose a two-stage method, i.e., collection and ranking, to recover PPI when PII entities have been masked.
+Different Settings. Some works also explore the risk of data leakage in novel settings. Wang et al. [555] study the probability of data extraction in fine-tuning settings and Bargav et al. [603, 604] extract the training data by comparing the output difference before and after the fine-tuning. Jiang et al. [605, 606, 607] propose to extract the private Retrieval-Augmented Generation (RAG) documents. Peng et al. [608] extract the private RAG documents by poisoning in the fine-tuning process. Nasr et al. [107] explore the potential risk of data extraction for the aligned production language models. Panda et al. [609] extract the fine-tuning secret data by poisoning the pertaining dataset. Lu et al. [610] propose to extract PII from an aligned model with model merging. Chen et al. [611] find that fine-tuning can recover the forgotten PIIs in pretraining data. Panchendrarajan et al. [612] propose to extract the whole private training data in the fine-tuning process. Rashid et al. [613] propose selective weight tampering to explore PPI leakage in Federated Language Models. Dentan et al. [614] extract data from layout-aware document understanding models like unimodal or bimodal models.
+Different Applications. Leveraging the abnormally high
+
+token probabilities, some works utilize the memorization of LLMs to extract the fingerprint or steganography [615]. Al-Kaswan et al. [616] explore memorization in large language models for code and find that code models memorize training data at a lower rate than natural language models. Nie et al. [617] utilize the token-level features derived from the identified characteristics to decode the PII. Lehman et al. [618] reveal the risk of Electronic Health Records leakage of LLMs. Diera et al. [619] conduct experiments to assess the PII leakage of fine-tuned BERT models and found that Differential Privacy (DP) has a negative effect when deployed in fine-tuning. Zhang et al. [620] propose data extraction attacks against text classification with transformers. Huang et al. [621] propose an evaluation tool, i.e. HCR, to assess the PPI leakage in Neural Code Completion Tools.
+
+$\nrightarrow$ Factor Assessment. Some work studies the factors of data extraction including decoding schemes, model sizes, prefix lengths, partial sequence leakages, and token positions [622, 623]. Yash et al. [624] explore the effects of prompt sensitivity and access to multiple checkpoints to extraction attacks. Staab et al. [625] construct a dataset consisting of real Reddit profiles to extract personal attributes. Xu et al. [626] conduct experiments to evaluate the factors of different suffix generation methods and different membership inference attacks in extraction performance. Karamolegkou et al. [627] evaluate the effect of model structure, data type, probing strategies, and metrics.
+
+Prompt Stealing Attacks. Given that crafting effective prompts requires significant engineering effort and can be considered valuable intellectual property (IP), promptstealing attacks aim to compromise this IP by reconstructing prompts from generated responses [556, 557, 558]. These generation effects are often used to attract prospective prospective buyers. Sha et al. [556] pioneer this approach by collecting a dataset and training classifiers to predict prompt parameters—such as whether the prompt is direct, role-based, or in-context. They then used a large language model (LLM) to reconstruct the prompt. Similarly, Zhang et al. [557] trained an LLM on output-prompt pairs to directly infer the original prompt, while Yang et al. [558] leveraged generation differences to refine surrogate prompts. However, recovering the original prompt solely from the output is challenging. Out of this, Zheng et al. [628] propose a timing-based side-channel method to infer the prompt during inference.
+
+# 6.1.2 Defensive Mechanisms in Deployment
+
+In Subsubsection 6.1.1, we analyzed various attack scenarios targeting individual LLM deployments. However, in real-world applications, defense mechanisms are not designed as isolated, one-to-one countermeasures against specific attacks. Instead, they follow fundamental security principles to establish a systematic defense framework, as illustrated in Figure 9. This framework integrates multiple layers of protection, ensuring resilience against a wide range of adversarial threats while maintaining model usability and efficiency.
+
+Input Preprocessing Defenses Input preprocessing serves as the first line of defense in LLM deployment, aiming to
+
+
+Defensive Mechanisms in Deployment
+Fig. 9: The overview of attacks in single LLM's deployment phase.
+
+detect and neutralize adversarial inputs before they reach the model.
+
+Attack Detection & Identification: Effective input filtering [629, 630] begins with attack detection [631], which identifies adversarial prompts through statistical [632], structural [633], or behavioral inconsistencies [634]. Gradient-based detection methods [635] leverage safety-critical gradient analysis and loss landscape exploration to uncover jailbreak prompts that manipulate LLM behavior. These approaches identify adversarial inputs [636, 637] by analyzing how small perturbations [638] affect model outputs, detecting highly sensitive or misaligned gradients that indicate targeted attacks. Perplexity-based methods [632, 632] measure the probability distribution of input sequences, flagging atypical or low-likelihood prompts as potential adversarial inputs. These techniques are particularly effective in detecting prompt injection and adversarial perturbations, where crafted prompts deviate significantly from natural language distributions.
+
+Beyond individual heuristics, universal detection frameworks [639] integrate multiple detection strategies to counter diverse attack vectors, including prompt injection [640], backdoor manipulations [641], and adversarial attacks [637]. These frameworks employ ensemble-based filtering mechanisms, combining gradient analysis [642], perplexity estimation [643], and syntactic evaluation for generalized attack resilience.
+
+Semantic & Behavioral Analysis: Attack detection alone is insufficient, as certain adversarial inputs may bypass traditional filtering mechanisms. Semantic [644] and behavioral analysis enhance input preprocessing by evaluating linguistic intent and model alignment. Self-examination techniques allow LLMs [645, 646] to assess whether they are being manipulated, leveraging auxiliary reasoning steps to detect deceptive prompts. Alignment-based verification [647] ensures that the model's responses remain consistent with its safety objectives [330], identifying inputs that subtly nudge the model toward policy violations or ethical misalignment. Intention analysis [648, 649] further refines input filtering by discerning subtle manipulations designed to bypass explicit security checks. Unlike token-level detection, which flags overtly adversarial inputs, intention-aware defenses analyze the semantic structure and purpose of the input to preemptively reject jailbreak attempts.
+
+Adversarial Defense & Mitigation: When detection and behavioral analysis fail to fully neutralize adversarial inputs, robustness-enhancing techniques [647] mitigate their effects by reducing model susceptibility to manipulation [334, 650]. Semantic smoothing [651, 652] techniques
+
+introduce controlled randomness into LLM responses, reducing the model's sensitivity to adversarial perturbations and preventing reliable jailbreak execution. By stabilizing decision boundaries [653], these methods enhance resistance against prompt manipulation strategies that exploit response predictability.
+
+Preemptive input transformations [654], such as back-translation [655] or paraphrasing, modify incoming queries [651] while preserving semantic intent, disrupting adversarial structures embedded within malicious prompts. Data augmentation [656] and adversarial training further strengthen model robustness by exposing LLMs to adversarial prompts during training, forcing them to learn invariances that reduce their vulnerability to real-world attacks.
+
+Output Filtering Mechanisms. Output filtering mechanisms [212, 657] serve as a critical safeguard in LLM deployment, ensuring that generated responses comply with safety constraints while preserving informativeness. Unlike input preprocessing, which aims to prevent adversarial prompts from reaching the model, output filtering mitigates harmful content post-generation. Existing approaches primarily follow three paradigms: rule-based constraints, generative adversarial filtering, and toxicity detection.
+
+Rule-based mechanisms [658] impose predefined constraints on model outputs, preventing the generation of harmful, unethical, or undesired content. Programmable guardrails [659] offer a structured framework where developers can enforce response filtering, topic restriction, and ethical alignment. These methods often integrate reinforcement learning from human feedback [155] or rule-based reward [660] modeling to refine output safety. While effective at handling explicit violations, static rule-based methods struggle with nuanced adversarial prompts and subtle misalignments.
+
+To address these limitations, generative adversarial filtering [661] leverages self-critique [662, 663], ensemble detection, and dynamic response evaluation [664]. Self-rectification mechanisms [663, 665] enable LLMs to critique their own outputs and refine responses through iterative refinement. Additionally, ensemble-based [666] moderation models aggregate predictions from multiple LLMs, improving robustness against circumvention techniques. Adaptive filtering frameworks [667] employ perplexity-based assessments and adversarial perturbation detection to flag responses deviating from expected linguistic patterns, enhancing their resilience against jailbreak attempts [668, 669] and toxic content injection.
+
+Toxicity detection [670, 671, 672] and content moderation [673, 674, 675, 676] further reinforce output safety by identifying and mitigating hate speech [677], misinformation, and other harmful content. Supervised finetuning adapts LLMs to recognize undesirable patterns, while classifier-based detection models [678] filter responses in real-time. Some approaches introduce debiasing strategies, such as controlled decoding [679, 680] and anti-expert guidance [681], to suppress toxic outputs without sacrificing response diversity. However, these methods face challenges in balancing false positives and false negatives, particularly in ambiguous or context-dependent cases.
+
+The effectiveness of output filtering hinges on its ability to balance strict control with linguistic flexibility, ensur
+
+ing that models remain both safe and practically useful. A hybrid approach combining rule-based safeguards, self-correcting mechanisms, and adaptive toxicity moderation is essential to achieving robust and scalable LLM deployment. Robust Prompt Engineering. Robust prompt engineering aims to enhance LLM safety by designing input prompts that resist adversarial manipulation [682], protect sensitive data, and mitigate harmful outputs—all [683] without modifying model parameters. These strategies act at the interaction level, offering lightweight and model-agnostic protection.
+
+Recent efforts have introduced prompt optimization techniques grounded in adversarial robustness, including embedding-space manipulation and defensive objective alignment. Methods such as Robust Prompt Optimization [684] and Prompt Adversarial Tuning generate transferable suffixes [668] or prefix [685] embeddings to guide model behavior [686] under attack [687], effectively lowering jailbreak success rates while preserving task performance. Similarly, goal prioritization frameworks [688] enforce inference-time objective consistency, dynamically resolving conflict between user instructions and safety constraints without requiring access to malicious samples. Complementary to these strategies, patch-based methods integrate interpretable suffixes or structured self-reminders [689] into prompts, reducing the model's susceptibility to coercive inputs through lightweight, modular defenses.
+
+Structural manipulation approaches [690] neutralize adversarial intent through prompt rewriting. Spotlighting [691] injects source-attribute signals to counter indirect prompt injection, while inverse prompt engineering [692] repurposes attack data to generate task-specific defensive prompts under the principle of least privilege.
+
+Privacy-preserving prompt [693] design introduces formal guarantees through differential privacy. Approaches like DP-Prompt [694] and stochastic gradient masking [695] reduce information leakage from prompts without harming performance. Desensitization and directional control of incontext representations offer additional privacy protections during prompt construction. Prompt engineering [579, 696] also helps mitigate societal risks. Chain-of-thought prompting and guided templates reduce gender bias [697] in reasoning tasks, while prompt learning [698] improves toxicity detection and generation control [699, 700], often surpassing specialized models in efficiency and generalization.
+
+Finally, systematic prompt optimization methods [701, 702] aim to generalize prompt robustness across tasks and domains. Techniques like BATPrompt [703] and StraGo [704] use adversarial simulation and strategic decomposition to refine prompts iteratively, improving both resilience and effectiveness under variable inputs.
+
+System-level Security Controls. System-level defenses [705] enhance LLM deployment by optimizing inference, enforcing alignment, isolating untrusted inputs, and securing the supply chain. Systems like Petals [706], Sarathi-Serve [707], and DistServe [708] restructure computation to improve throughput and latency, while TriForce [709], Medusa [710] MagicDec [711] accelerate generation via speculative decoding and structural compression. Parallel frameworks such as DeepSpeed-FastGen [712] and SpecExec [713] further boost
+
+efficiency with minimal overhead.
+
+Runtime alignment methods [714] adapt model behavior through cross-model guidance or token-level reward modeling. Systems such as SelfDefend [715] and Gradient Cuff [716] detect unsafe generation by monitoring agreement across models or loss landscapes, while Spotlighting [691] inserts provenance signals to mitigate indirect prompt injection.
+
+Access isolation is achieved through policy enforcement [717] and system wrappers [688]. At the supply level, tools like MalHug [718] identify poisoned models, while system audits reveal sandbox and plugin vulnerabilities, highlighting the need for end-to-end secure deployment.
+
+LLM-based guard models utilize lightweight LLMs like Llama Guard [330], Aegis Guard [719, 720], WildGuard [721], and ShieldGemma [722] to moderate both the input and output of the victim LLMs. However, they are purely classifiers. To solve this problem, the first reasoning-based guard model named GuardReasoner [723] is proposed to improve the performance, explainability, and generalization ability via learning to reason. It brings new opportunities for the safety of large-scale reasoning models [724].
+
+# 6.1.3 Evaluation and Benchmarks in Deployment
+
+To assess the reliability and safety of LLMs after deployment, evaluation efforts focus on several key dimensions and risk types, as illustrated in Figure 10. These dimensions guide the design of systematic benchmarks and metrics tailored for real-world deployment settings.
+
+
+Fig. 10: The overview of evaluation and benchmarks in single LLM's deployment phase.
+
+Robustness Evaluation. To systematically assess the relia-
+
+TABLE 7: Summary of LLM robustness benchmarks at the deployment stage.
+
+| Attacks | Method | OS | Year | Strategy | Setting | Datasets | Target Models | Metrics |
| Model Extraction | Carlini et al. [503] | Yes | 2024 | Binary Search | Black-box | None | GPTs, LLaMA, Pythia,ada, cabbage | Query&TokenCost, MSE, RMS |
| Finlayson et al. [504] | No | 2024 | Softmax Bottleneck | Black-box | None | Pythia, GPT-3.5 | Query Cost |
| Zanella et al. [505] | No | 2024 | Matrix Operations | Grey-box | SST-2, MNLI, AGNews | BERTs, XLNet | Query Cost, Acc,Agreement |
| Horwitz et al. [506] | Yes | 2024 | Spectral DeTuning | White-box | LoWRA | ViT, SD, Mistral | MSWE, SEM |
| Membership Inference | MIN-K% PROB [509] | Yes | 2023 | Probabilities | Black-box | Wikipedia | LLaMAs, Pythia, NeoX,OPT | TPR, FPR, ROC,AUC |
| MIN-K%++ [510] | Yes | 2022 | Local Maxima | Black-box | WikiMIA, MIMIR | Pythia, GPT-NeoXLLaMA, OPT, Mamba | AUROC, TPR, FPR |
| Blind [511] | Yes | 2024 | Threshold | Black-box | 8 sets | GPT-3, OpenLLaMA | AUC ROC |
| LLM-DI [512] | Yes | 2024 | Aggregation | Black-box | PILE | Pythias | AUC, p-values |
| DE-COP [513] | Yes | 2024 | Paraphrases | Black-box | arXiv:Tection, BookTection | Mistral, Mixtral, LLaMA, GPTs, Claude | AUC |
| Recall [514] | Yes | 2024 | Log-Likelihoods | Black-box | WikiMIA, MIMIR | Pythia, GPT-NeoXLLaMA, OPT, Mamba | AUC, TPR@FPR |
| Noisy [515] | No | 2024 | Embedding NGBRs | Gray-box | OpenWebText,Wikipedia | GPT-2 | TPR, FPR, AUC |
| SMIA [516] | No | 2024 | Perturbation | Gray-box | Wikipedia, FAN | Pythia, Pythia-Deduped, GPT-Neos | AUC-ROC, TPR, FPR |
| FEATAGG [517] | No | 2024 | Feature Aggregation | Black-box | ProjectGutenberg,ArXiv | OpenLLaMA | TPR@FPR, AUC |
| RAG-MIA [520] | No | 2024 | Direct Asking | Black-box | HealthCareMagic,Enron | flan, llama, mistral | TPR@FPR, AUC-ROC |
| Jailbreak | GCG [260] | Yes | 2023 | Gradient-based | White-box | Vicuna, LLaMA-2 | AdvBench | ASR, Loss |
| AmpleGCG [530] | Yes | 2024 | Hybrid-based | White-box | Vicuna, Llama-2,Mis-tral,GPTs | AdvBench | ASR, US, Diver-sity, Time |
| I-GCG [531] | Yes | 2024 | Gradient-based | White-box | AdvBench,HarmBench | VICUNA, GUANACOLLAMA, MISTRAL | ASR |
| MA-GCG [532] | Yes | 2024 | Gradient-based | White-box | AdvBench | Vicuna, Mistral | ASR, Time |
| A-GCG [533] | Yes | 2024 | Gradient-based | White-box | AdvBench | Llama2, Vicuna | ASR, Acc |
| AutoDAN-A [534] | Yes | 2023 | LLM-based | Black-box | AdvBench | Vicuna, Mistral | ASR, Recheck,PPL |
| AutoDAN-B [535] | Yes | 2023 | Gradient-based | White-box | AdvBench | Vicuna, Guanaco, Pythia | ASR, Recheck |
| PAIR [261] | Yes | 2023 | LLM-based | Black-box | JailbreakBench | Vicuna, Llama-2, GPTs,Claudes,Gemini | ASR, QPS |
| ToA [536] | Yes | 2023 | LLM-based | Black-box | AdvBench, Harm123 | Vicuna, Llama-2, PaLM-2,GPTs, Claude3, Gemini | GPT4-MetricHuman-Judge |
| PAL [537] | Yes | 2024 | LLM-based | Black-box | AdvBench | Llama-2, GPT-3.5 | ASR, Manual Labeling |
| Masterkey [538] | No | 2023 | Rephrasing | Black-box | AdvBench, Harm123 | GPTs, Bing, Bard | ASR, QSR |
| AutoDAN-Turbo [539] | Yes | 2024 | LLM-based | Black-box | Harmbench | Llama-2, Gemma, GPT-4,Gemini | ASR, StrongRE-JECT |
| FlipAttack [540] | Yes | 2025 | Rephrasing | Black-box | AdvBench, StrongRE-JECT | GPTs, Claude 3.5 Sonnet, Llama 3.1 405B, Mixtral 8x22B | ASR |
| Geneshift [541] | Yes | 2025 | LLM-based | Black-box | AdvBench | GPTs | ASR |
| Prompt Injection | IPP [542] | Yes | 2022 | Handcraft | Black-box | OpenAI Examples | text-davinci | ASR |
| Greshake et al. [543] | Yes | 2023 | Data Poisoning | Black-box | None | text-davinci, GPT-4 | None |
| HOUYI [544] | Yes | 2023 | Components Asmbl | Black-box | Five Queries | SUPERTOOLS | Manual |
| Yan et al. [130] | Yes | 2023 | Poisoning | Black-box | Several Cases | Alpaca | Ngt, Pst, Ocrc |
| TT [545] | No | 2023 | Game | Black-box | Tensor Trust | GPTs, Claudes, PaLM, LLaMAs | Robustness Rate |
| JudgeDeceiver [546] | Yes | 2024 | Gradient-based | White-box | MT-Bench, LLMBarMRPC, Jfleg, HSOL,RTE, SST2, SMS | Mistral, Openchat, Llamas | ACC, ASR, PACKEY-E, LM-E |
| AUPI [547] | Yes | 2024 | Gradient-based | White-box | MRPC, Jfleg, HSOL,RTE, SST2, SMS | Llama2 | ASR |
| AUTOHIJACKER [548] | No | 2024 | LLM-based | Black-box | AgentDojo, OPI | Llama, Command-R,GPTs | ASR |
| Data Extraction | zlib [108] | Yes | 2020 | Generate & Inference | Black-box | Top-n, Temperature, Internet | GPT-2 | 6 metrics |
| AutoSklearn [549] | No | 2023 | Greedy, Contrastive, Beam decoding | Black-box | Pile | GPT-Neo | Precision, Recall,R@FPR |
| DECOM [550] | No | 2024 | Decomposition | Black-box | NYT, WSJ | Frontiers | TRM, EMP,BITAP |
| Context [551] | No | 2022 | Context, Zero-shot,Few-shot | Black-box | Enron Corpus | GPT-Neo | Acc |
| ETHICIST [552] | Yes | 2023 | Prompt Tuning | Gray-box | LM-Extraction | GPT-Neo | Recall |
| Pli-compass [553] | No | 2024 | Grounding | Black-box | Enron email | GPT-J | Extraction Rate |
| DSP [554] | No | 2024 | Dynamic Soft Prompting | Black-box | LMEB, The Stack | GPT-Neo, Pythia, Star-CoderBase | EER, FER, PPL |
| PWB [555] | Yes | 2024 | Gradient-based | White-box | Pile | Pythia, Llama | Precision, AUC,TPR |
| Prompt Stealing | Sha et al. [556] | No | 2024 | LLM-based | Black-box | RetrievalQA,AlpacaGPT4 | ChatGPT, LLaMA | Acc, Precision, Recall, AUC |
| output2prompt [557] | Yes | 2024 | LLM-based | Black-box | 3 User & 3 SystemPrompts | Llamas, GPTs | BLEU, CS, Preci-sion, Recall |
| PRSA [558] | No | 2024 | Output Difference | Black-box | Category18 | GPTs | BLEU, FastKAS-SIM, JS |
",
+ "image_path": "1b920951359c81b1de588e6f741d867a8d8080726518d9c47818f949e3c7423e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 45,
+ 25,
+ 258,
+ 35
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 25,
+ 258,
+ 35
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 25,
+ 258,
+ 35
+ ],
+ "type": "text",
+ "content": "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 555,
+ 26,
+ 564,
+ 34
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 555,
+ 26,
+ 564,
+ 34
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 555,
+ 26,
+ 564,
+ 34
+ ],
+ "type": "text",
+ "content": "20"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 19
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 52,
+ 41,
+ 300,
+ 88
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 41,
+ 300,
+ 88
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 41,
+ 300,
+ 88
+ ],
+ "type": "text",
+ "content": "ECLIPSE [588] uses an LLM as a suffix generator and optimizer. PAL [537] proposes an online proxy model (which is used for adversarial prompt generation) training pipeline."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 41,
+ 88,
+ 301,
+ 170
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 41,
+ 88,
+ 301,
+ 170
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 41,
+ 88,
+ 301,
+ 170
+ ],
+ "type": "inline_equation",
+ "content": "\\Rightarrow"
+ },
+ {
+ "bbox": [
+ 41,
+ 88,
+ 301,
+ 170
+ ],
+ "type": "text",
+ "content": " Others. EnJa [589] proposes to ensemble prompt and token-level attack methods via a template-based connector. AmpleGCG [530] first collects lots of successful suffixes and then trains the generative model to generate a specific suffix for a given unsafe prompt. Zhao et al. [590] targets the scenario where the decoding process of target LLM is assisted with smaller models' guidance."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 41,
+ 175,
+ 301,
+ 256
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 41,
+ 175,
+ 301,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 41,
+ 175,
+ 301,
+ 256
+ ],
+ "type": "text",
+ "content": "Prompt Injection Attacks. Prompt injection is a vulnerability where an attacker manipulates the input prompts of LLMs to force them to generate a specific output, which is usually out of the range for normal use (e.g., goal hijacking and prompt leaking [542]), often by injecting malicious text or commands into the input field. Attackers can employ a variety of techniques to carry out such attacks."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 41,
+ 261,
+ 301,
+ 747
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 8,
+ "blocks": [
+ {
+ "bbox": [
+ 41,
+ 261,
+ 301,
+ 469
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 41,
+ 261,
+ 301,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 41,
+ 261,
+ 301,
+ 469
+ ],
+ "type": "inline_equation",
+ "content": "\\Rightarrow"
+ },
+ {
+ "bbox": [
+ 41,
+ 261,
+ 301,
+ 469
+ ],
+ "type": "text",
+ "content": " Direct Prompt Injection. Perez et al. [542] directly inject handcrafted adversarial prompts into inputs to misalign the language model. HOUYI [544] proposes an injection generation framework which includes three components. Yan et al. [130] utilize LLMs to generate diverse trigger instructions that implicitly capture the characteristics of trigger scenarios. TENSOR TRUST leverages the TENSOR TRUST web game to generate a large-scale dataset and benchmark [545]. AUPI [547] adopts a gradient-based optimization method, specifically, a momentum-enhanced optimization algorithm, to generate universal prompt injection data. Upadhayay et al. [591] argue that LLMs suffer from cognitive overload and propose to use in-context learning to jailbreak LLMs through deliberately designed prompts that induce cognitive overload. Kwon et al. [592] circumvent security policies by substituting sensitive words—likely to be rejected by the language model—with mathematical functions."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 41,
+ 469,
+ 301,
+ 700
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 41,
+ 469,
+ 301,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 41,
+ 469,
+ 301,
+ 700
+ ],
+ "type": "inline_equation",
+ "content": "\\nLeftrightarrow"
+ },
+ {
+ "bbox": [
+ 41,
+ 469,
+ 301,
+ 700
+ ],
+ "type": "text",
+ "content": " Indirect Prompt Injection. Greshake et al. [543] propose to indirectly inject prompts into the data that are likely to be retrieved. Bagdasaryan et al. [593] design a prompt injection attack against multi-modal LLMs, by generating an adversarial perturbation corresponding to the prompt and blending it into an image or audio recording. Neural Exec [594] designs a multi-stage preprocessing pipeline for cases like Retrieval-Augmented Generation (RAG)-based applications. PoisonedAlign [595] boosts the success of prompt injection attacks by strategically creating poisoned alignment samples in the LLM's alignment process. TPIA [596] crafts non-functional perturbations that contain malicious information and inserts them into the victim's code context by spreading them into potentially used dependencies like packages or RAG's knowledge base. F2A [597] proposes to use feign security detection agents to bypass the defense mechanism of LLMs. AUTOHIJACKER [548] uses a batch-based optimization framework to handle sparse feedback and leverages a trainable memory to enable effective generation."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 41,
+ 700,
+ 301,
+ 747
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 41,
+ 700,
+ 301,
+ 747
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 41,
+ 700,
+ 301,
+ 747
+ ],
+ "type": "text",
+ "content": "Different Settings. JudgeDeceiver uses gradient-based optimization to inject LLM-as-a-Judge scenarios [546]. Pedro et al. [598] study the risk of injections targeting web applications based on the Langchain framework. Lee et"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 315,
+ 41,
+ 564,
+ 124
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 41,
+ 564,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 41,
+ 564,
+ 124
+ ],
+ "type": "text",
+ "content": "al. [599] propose a human-AI collaborative framework to explore the potential of prompt injection against federated military LLMs. PROMPT INFECTION [600] proposes to make malicious prompts self-replicate across interconnected agents in multi-agent systems. Zhang et al. [601] explore the risk of prompt injection in LLM-integrated systems like LLM-integrated mobile robotic systems."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 308,
+ 129,
+ 565,
+ 186
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 129,
+ 565,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 129,
+ 565,
+ 186
+ ],
+ "type": "text",
+ "content": "Data Extraction Attacks. Data extraction attacks try to figure out the personally identifiable information (PII) that is used to train the LLMs [108]. It starts from sufficient-length prefixes to perform extraction and additional measures to determine if extracted texts are valid."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 306,
+ 192,
+ 565,
+ 747
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 14,
+ "blocks": [
+ {
+ "bbox": [
+ 307,
+ 192,
+ 565,
+ 479
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 192,
+ 565,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 192,
+ 565,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "\\nLeftrightarrow"
+ },
+ {
+ "bbox": [
+ 307,
+ 192,
+ 565,
+ 479
+ ],
+ "type": "text",
+ "content": " Methods. In the beginning work [108], the proposed extraction process contains two stages \"generate-then-rank\": sampling potentially memorized examples and membership inference. It proposes a temperature-decaying method to sample more diverse examples and use surrogate models to infer the membership. After that, Al-Kaswan et al. [549] propose using greedy, contrastive, and beam decoding strategies to generate examples and use a classifier to infer the membership. Su et al. [550] propose an instruction decomposition technique to extract fragments of training data gradually. Huang et al. [551] extensively explore the effect of context, zero-shot, and few-shot methods in extracting the personal email address. ETHICIST proposes a smoothing loss and a calibrated confidence estimation method to extract the suffix and measure the confidence [552]. Nakka et al. [553] improves the extraction performance by grounding the prefix of the manually constructed extraction prompt with in-domain data. Wang et al. [554] propose to train a transformer-based generator to produce dynamic, prefix-dependent soft prompts. Ozdayi et al. [105] introduce an approach that uses prompt tuning to control the extraction rates of memorized content. Meng et al. [602] propose a two-stage method, i.e., collection and ranking, to recover PPI when PII entities have been masked."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 307,
+ 481,
+ 565,
+ 734
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 481,
+ 565,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 481,
+ 565,
+ 734
+ ],
+ "type": "text",
+ "content": "Different Settings. Some works also explore the risk of data leakage in novel settings. Wang et al. [555] study the probability of data extraction in fine-tuning settings and Bargav et al. [603, 604] extract the training data by comparing the output difference before and after the fine-tuning. Jiang et al. [605, 606, 607] propose to extract the private Retrieval-Augmented Generation (RAG) documents. Peng et al. [608] extract the private RAG documents by poisoning in the fine-tuning process. Nasr et al. [107] explore the potential risk of data extraction for the aligned production language models. Panda et al. [609] extract the fine-tuning secret data by poisoning the pertaining dataset. Lu et al. [610] propose to extract PII from an aligned model with model merging. Chen et al. [611] find that fine-tuning can recover the forgotten PIIs in pretraining data. Panchendrarajan et al. [612] propose to extract the whole private training data in the fine-tuning process. Rashid et al. [613] propose selective weight tampering to explore PPI leakage in Federated Language Models. Dentan et al. [614] extract data from layout-aware document understanding models like unimodal or bimodal models."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 306,
+ 734,
+ 564,
+ 747
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 734,
+ 564,
+ 747
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 734,
+ 564,
+ 747
+ ],
+ "type": "text",
+ "content": "Different Applications. Leveraging the abnormally high"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "sub_type": "text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 45,
+ 25,
+ 258,
+ 35
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 25,
+ 258,
+ 35
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 25,
+ 258,
+ 35
+ ],
+ "type": "text",
+ "content": "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 555,
+ 26,
+ 563,
+ 34
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 555,
+ 26,
+ 563,
+ 34
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 555,
+ 26,
+ 563,
+ 34
+ ],
+ "type": "text",
+ "content": "21"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 20
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 52,
+ 42,
+ 301,
+ 227
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 42,
+ 301,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 42,
+ 301,
+ 227
+ ],
+ "type": "text",
+ "content": "token probabilities, some works utilize the memorization of LLMs to extract the fingerprint or steganography [615]. Al-Kaswan et al. [616] explore memorization in large language models for code and find that code models memorize training data at a lower rate than natural language models. Nie et al. [617] utilize the token-level features derived from the identified characteristics to decode the PII. Lehman et al. [618] reveal the risk of Electronic Health Records leakage of LLMs. Diera et al. [619] conduct experiments to assess the PII leakage of fine-tuned BERT models and found that Differential Privacy (DP) has a negative effect when deployed in fine-tuning. Zhang et al. [620] propose data extraction attacks against text classification with transformers. Huang et al. [621] propose an evaluation tool, i.e. HCR, to assess the PPI leakage in Neural Code Completion Tools."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 42,
+ 228,
+ 301,
+ 376
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 42,
+ 228,
+ 301,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 42,
+ 228,
+ 301,
+ 376
+ ],
+ "type": "inline_equation",
+ "content": "\\nrightarrow"
+ },
+ {
+ "bbox": [
+ 42,
+ 228,
+ 301,
+ 376
+ ],
+ "type": "text",
+ "content": " Factor Assessment. Some work studies the factors of data extraction including decoding schemes, model sizes, prefix lengths, partial sequence leakages, and token positions [622, 623]. Yash et al. [624] explore the effects of prompt sensitivity and access to multiple checkpoints to extraction attacks. Staab et al. [625] construct a dataset consisting of real Reddit profiles to extract personal attributes. Xu et al. [626] conduct experiments to evaluate the factors of different suffix generation methods and different membership inference attacks in extraction performance. Karamolegkou et al. [627] evaluate the effect of model structure, data type, probing strategies, and metrics."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 44,
+ 378,
+ 301,
+ 586
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 378,
+ 301,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 378,
+ 301,
+ 586
+ ],
+ "type": "text",
+ "content": "Prompt Stealing Attacks. Given that crafting effective prompts requires significant engineering effort and can be considered valuable intellectual property (IP), promptstealing attacks aim to compromise this IP by reconstructing prompts from generated responses [556, 557, 558]. These generation effects are often used to attract prospective prospective buyers. Sha et al. [556] pioneer this approach by collecting a dataset and training classifiers to predict prompt parameters—such as whether the prompt is direct, role-based, or in-context. They then used a large language model (LLM) to reconstruct the prompt. Similarly, Zhang et al. [557] trained an LLM on output-prompt pairs to directly infer the original prompt, while Yang et al. [558] leveraged generation differences to refine surrogate prompts. However, recovering the original prompt solely from the output is challenging. Out of this, Zheng et al. [628] propose a timing-based side-channel method to infer the prompt during inference."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 45,
+ 594,
+ 241,
+ 606
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 594,
+ 241,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 594,
+ 241,
+ 606
+ ],
+ "type": "text",
+ "content": "6.1.2 Defensive Mechanisms in Deployment"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 44,
+ 608,
+ 300,
+ 723
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 608,
+ 300,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 608,
+ 300,
+ 723
+ ],
+ "type": "text",
+ "content": "In Subsubsection 6.1.1, we analyzed various attack scenarios targeting individual LLM deployments. However, in real-world applications, defense mechanisms are not designed as isolated, one-to-one countermeasures against specific attacks. Instead, they follow fundamental security principles to establish a systematic defense framework, as illustrated in Figure 9. This framework integrates multiple layers of protection, ensuring resilience against a wide range of adversarial threats while maintaining model usability and efficiency."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 44,
+ 723,
+ 301,
+ 748
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 723,
+ 301,
+ 748
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 723,
+ 301,
+ 748
+ ],
+ "type": "text",
+ "content": "Input Preprocessing Defenses Input preprocessing serves as the first line of defense in LLM deployment, aiming to"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 310,
+ 57,
+ 560,
+ 126
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 310,
+ 46,
+ 467,
+ 57
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 310,
+ 46,
+ 467,
+ 57
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 310,
+ 46,
+ 467,
+ 57
+ ],
+ "type": "text",
+ "content": "Defensive Mechanisms in Deployment"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 310,
+ 57,
+ 560,
+ 126
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 310,
+ 57,
+ 560,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 310,
+ 57,
+ 560,
+ 126
+ ],
+ "type": "image",
+ "image_path": "f4ce229cb80a8c96656ded2aa655c39cbe884dd49ac1e9e0e1650264348f5451.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 309,
+ 133,
+ 564,
+ 158
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 309,
+ 133,
+ 564,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 133,
+ 564,
+ 158
+ ],
+ "type": "text",
+ "content": "Fig. 9: The overview of attacks in single LLM's deployment phase."
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 308,
+ 178,
+ 564,
+ 201
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 178,
+ 564,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 178,
+ 564,
+ 201
+ ],
+ "type": "text",
+ "content": "detect and neutralize adversarial inputs before they reach the model."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 307,
+ 202,
+ 564,
+ 399
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 202,
+ 564,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 202,
+ 564,
+ 399
+ ],
+ "type": "text",
+ "content": "Attack Detection & Identification: Effective input filtering [629, 630] begins with attack detection [631], which identifies adversarial prompts through statistical [632], structural [633], or behavioral inconsistencies [634]. Gradient-based detection methods [635] leverage safety-critical gradient analysis and loss landscape exploration to uncover jailbreak prompts that manipulate LLM behavior. These approaches identify adversarial inputs [636, 637] by analyzing how small perturbations [638] affect model outputs, detecting highly sensitive or misaligned gradients that indicate targeted attacks. Perplexity-based methods [632, 632] measure the probability distribution of input sequences, flagging atypical or low-likelihood prompts as potential adversarial inputs. These techniques are particularly effective in detecting prompt injection and adversarial perturbations, where crafted prompts deviate significantly from natural language distributions."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 307,
+ 399,
+ 564,
+ 491
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 399,
+ 564,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 399,
+ 564,
+ 491
+ ],
+ "type": "text",
+ "content": "Beyond individual heuristics, universal detection frameworks [639] integrate multiple detection strategies to counter diverse attack vectors, including prompt injection [640], backdoor manipulations [641], and adversarial attacks [637]. These frameworks employ ensemble-based filtering mechanisms, combining gradient analysis [642], perplexity estimation [643], and syntactic evaluation for generalized attack resilience."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 307,
+ 492,
+ 564,
+ 689
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 492,
+ 564,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 492,
+ 564,
+ 689
+ ],
+ "type": "text",
+ "content": "Semantic & Behavioral Analysis: Attack detection alone is insufficient, as certain adversarial inputs may bypass traditional filtering mechanisms. Semantic [644] and behavioral analysis enhance input preprocessing by evaluating linguistic intent and model alignment. Self-examination techniques allow LLMs [645, 646] to assess whether they are being manipulated, leveraging auxiliary reasoning steps to detect deceptive prompts. Alignment-based verification [647] ensures that the model's responses remain consistent with its safety objectives [330], identifying inputs that subtly nudge the model toward policy violations or ethical misalignment. Intention analysis [648, 649] further refines input filtering by discerning subtle manipulations designed to bypass explicit security checks. Unlike token-level detection, which flags overtly adversarial inputs, intention-aware defenses analyze the semantic structure and purpose of the input to preemptively reject jailbreak attempts."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 308,
+ 689,
+ 564,
+ 748
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 689,
+ 564,
+ 748
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 689,
+ 564,
+ 748
+ ],
+ "type": "text",
+ "content": "Adversarial Defense & Mitigation: When detection and behavioral analysis fail to fully neutralize adversarial inputs, robustness-enhancing techniques [647] mitigate their effects by reducing model susceptibility to manipulation [334, 650]. Semantic smoothing [651, 652] techniques"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 45,
+ 26,
+ 258,
+ 35
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 26,
+ 258,
+ 35
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 26,
+ 258,
+ 35
+ ],
+ "type": "text",
+ "content": "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 555,
+ 26,
+ 564,
+ 34
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 555,
+ 26,
+ 564,
+ 34
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 555,
+ 26,
+ 564,
+ 34
+ ],
+ "type": "text",
+ "content": "22"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 21
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 44,
+ 42,
+ 301,
+ 112
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 42,
+ 301,
+ 112
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 42,
+ 301,
+ 112
+ ],
+ "type": "text",
+ "content": "introduce controlled randomness into LLM responses, reducing the model's sensitivity to adversarial perturbations and preventing reliable jailbreak execution. By stabilizing decision boundaries [653], these methods enhance resistance against prompt manipulation strategies that exploit response predictability."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 44,
+ 112,
+ 301,
+ 204
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 112,
+ 301,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 112,
+ 301,
+ 204
+ ],
+ "type": "text",
+ "content": "Preemptive input transformations [654], such as back-translation [655] or paraphrasing, modify incoming queries [651] while preserving semantic intent, disrupting adversarial structures embedded within malicious prompts. Data augmentation [656] and adversarial training further strengthen model robustness by exposing LLMs to adversarial prompts during training, forcing them to learn invariances that reduce their vulnerability to real-world attacks."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 44,
+ 204,
+ 301,
+ 308
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 204,
+ 301,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 204,
+ 301,
+ 308
+ ],
+ "type": "text",
+ "content": "Output Filtering Mechanisms. Output filtering mechanisms [212, 657] serve as a critical safeguard in LLM deployment, ensuring that generated responses comply with safety constraints while preserving informativeness. Unlike input preprocessing, which aims to prevent adversarial prompts from reaching the model, output filtering mitigates harmful content post-generation. Existing approaches primarily follow three paradigms: rule-based constraints, generative adversarial filtering, and toxicity detection."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 47,
+ 308,
+ 301,
+ 434
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 308,
+ 301,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 308,
+ 301,
+ 434
+ ],
+ "type": "text",
+ "content": "Rule-based mechanisms [658] impose predefined constraints on model outputs, preventing the generation of harmful, unethical, or undesired content. Programmable guardrails [659] offer a structured framework where developers can enforce response filtering, topic restriction, and ethical alignment. These methods often integrate reinforcement learning from human feedback [155] or rule-based reward [660] modeling to refine output safety. While effective at handling explicit violations, static rule-based methods struggle with nuanced adversarial prompts and subtle misalignments."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 44,
+ 435,
+ 301,
+ 584
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 435,
+ 301,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 435,
+ 301,
+ 584
+ ],
+ "type": "text",
+ "content": "To address these limitations, generative adversarial filtering [661] leverages self-critique [662, 663], ensemble detection, and dynamic response evaluation [664]. Self-rectification mechanisms [663, 665] enable LLMs to critique their own outputs and refine responses through iterative refinement. Additionally, ensemble-based [666] moderation models aggregate predictions from multiple LLMs, improving robustness against circumvention techniques. Adaptive filtering frameworks [667] employ perplexity-based assessments and adversarial perturbation detection to flag responses deviating from expected linguistic patterns, enhancing their resilience against jailbreak attempts [668, 669] and toxic content injection."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 44,
+ 585,
+ 301,
+ 723
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 585,
+ 301,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 585,
+ 301,
+ 723
+ ],
+ "type": "text",
+ "content": "Toxicity detection [670, 671, 672] and content moderation [673, 674, 675, 676] further reinforce output safety by identifying and mitigating hate speech [677], misinformation, and other harmful content. Supervised finetuning adapts LLMs to recognize undesirable patterns, while classifier-based detection models [678] filter responses in real-time. Some approaches introduce debiasing strategies, such as controlled decoding [679, 680] and anti-expert guidance [681], to suppress toxic outputs without sacrificing response diversity. However, these methods face challenges in balancing false positives and false negatives, particularly in ambiguous or context-dependent cases."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 45,
+ 723,
+ 301,
+ 746
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 723,
+ 301,
+ 746
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 723,
+ 301,
+ 746
+ ],
+ "type": "text",
+ "content": "The effectiveness of output filtering hinges on its ability to balance strict control with linguistic flexibility, ensur"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 307,
+ 42,
+ 564,
+ 168
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 42,
+ 564,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 42,
+ 564,
+ 168
+ ],
+ "type": "text",
+ "content": "ing that models remain both safe and practically useful. A hybrid approach combining rule-based safeguards, self-correcting mechanisms, and adaptive toxicity moderation is essential to achieving robust and scalable LLM deployment. Robust Prompt Engineering. Robust prompt engineering aims to enhance LLM safety by designing input prompts that resist adversarial manipulation [682], protect sensitive data, and mitigate harmful outputs—all [683] without modifying model parameters. These strategies act at the interaction level, offering lightweight and model-agnostic protection."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 308,
+ 169,
+ 565,
+ 365
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 169,
+ 565,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 169,
+ 565,
+ 365
+ ],
+ "type": "text",
+ "content": "Recent efforts have introduced prompt optimization techniques grounded in adversarial robustness, including embedding-space manipulation and defensive objective alignment. Methods such as Robust Prompt Optimization [684] and Prompt Adversarial Tuning generate transferable suffixes [668] or prefix [685] embeddings to guide model behavior [686] under attack [687], effectively lowering jailbreak success rates while preserving task performance. Similarly, goal prioritization frameworks [688] enforce inference-time objective consistency, dynamically resolving conflict between user instructions and safety constraints without requiring access to malicious samples. Complementary to these strategies, patch-based methods integrate interpretable suffixes or structured self-reminders [689] into prompts, reducing the model's susceptibility to coercive inputs through lightweight, modular defenses."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 308,
+ 365,
+ 565,
+ 435
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 365,
+ 565,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 365,
+ 565,
+ 435
+ ],
+ "type": "text",
+ "content": "Structural manipulation approaches [690] neutralize adversarial intent through prompt rewriting. Spotlighting [691] injects source-attribute signals to counter indirect prompt injection, while inverse prompt engineering [692] repurposes attack data to generate task-specific defensive prompts under the principle of least privilege."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 308,
+ 435,
+ 565,
+ 573
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 435,
+ 565,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 435,
+ 565,
+ 573
+ ],
+ "type": "text",
+ "content": "Privacy-preserving prompt [693] design introduces formal guarantees through differential privacy. Approaches like DP-Prompt [694] and stochastic gradient masking [695] reduce information leakage from prompts without harming performance. Desensitization and directional control of incontext representations offer additional privacy protections during prompt construction. Prompt engineering [579, 696] also helps mitigate societal risks. Chain-of-thought prompting and guided templates reduce gender bias [697] in reasoning tasks, while prompt learning [698] improves toxicity detection and generation control [699, 700], often surpassing specialized models in efficiency and generalization."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 308,
+ 573,
+ 564,
+ 643
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 573,
+ 564,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 573,
+ 564,
+ 643
+ ],
+ "type": "text",
+ "content": "Finally, systematic prompt optimization methods [701, 702] aim to generalize prompt robustness across tasks and domains. Techniques like BATPrompt [703] and StraGo [704] use adversarial simulation and strategic decomposition to refine prompts iteratively, improving both resilience and effectiveness under variable inputs."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 308,
+ 643,
+ 565,
+ 746
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 643,
+ 565,
+ 746
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 643,
+ 565,
+ 746
+ ],
+ "type": "text",
+ "content": "System-level Security Controls. System-level defenses [705] enhance LLM deployment by optimizing inference, enforcing alignment, isolating untrusted inputs, and securing the supply chain. Systems like Petals [706], Sarathi-Serve [707], and DistServe [708] restructure computation to improve throughput and latency, while TriForce [709], Medusa [710] MagicDec [711] accelerate generation via speculative decoding and structural compression. Parallel frameworks such as DeepSpeed-FastGen [712] and SpecExec [713] further boost"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 45,
+ 25,
+ 258,
+ 35
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 25,
+ 258,
+ 35
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 25,
+ 258,
+ 35
+ ],
+ "type": "text",
+ "content": "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 555,
+ 26,
+ 564,
+ 34
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 555,
+ 26,
+ 564,
+ 34
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 555,
+ 26,
+ 564,
+ 34
+ ],
+ "type": "text",
+ "content": "23"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 22
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 44,
+ 42,
+ 192,
+ 53
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 42,
+ 192,
+ 53
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 42,
+ 192,
+ 53
+ ],
+ "type": "text",
+ "content": "efficiency with minimal overhead."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 44,
+ 53,
+ 300,
+ 135
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 53,
+ 300,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 53,
+ 300,
+ 135
+ ],
+ "type": "text",
+ "content": "Runtime alignment methods [714] adapt model behavior through cross-model guidance or token-level reward modeling. Systems such as SelfDefend [715] and Gradient Cuff [716] detect unsafe generation by monitoring agreement across models or loss landscapes, while Spotlighting [691] inserts provenance signals to mitigate indirect prompt injection."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 44,
+ 135,
+ 301,
+ 194
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 135,
+ 301,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 135,
+ 301,
+ 194
+ ],
+ "type": "text",
+ "content": "Access isolation is achieved through policy enforcement [717] and system wrappers [688]. At the supply level, tools like MalHug [718] identify poisoned models, while system audits reveal sandbox and plugin vulnerabilities, highlighting the need for end-to-end secure deployment."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 44,
+ 194,
+ 300,
+ 299
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 194,
+ 300,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 194,
+ 300,
+ 299
+ ],
+ "type": "text",
+ "content": "LLM-based guard models utilize lightweight LLMs like Llama Guard [330], Aegis Guard [719, 720], WildGuard [721], and ShieldGemma [722] to moderate both the input and output of the victim LLMs. However, they are purely classifiers. To solve this problem, the first reasoning-based guard model named GuardReasoner [723] is proposed to improve the performance, explainability, and generalization ability via learning to reason. It brings new opportunities for the safety of large-scale reasoning models [724]."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 45,
+ 311,
+ 261,
+ 323
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 311,
+ 261,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 311,
+ 261,
+ 323
+ ],
+ "type": "text",
+ "content": "6.1.3 Evaluation and Benchmarks in Deployment"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 44,
+ 326,
+ 301,
+ 385
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 326,
+ 301,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 326,
+ 301,
+ 385
+ ],
+ "type": "text",
+ "content": "To assess the reliability and safety of LLMs after deployment, evaluation efforts focus on several key dimensions and risk types, as illustrated in Figure 10. These dimensions guide the design of systematic benchmarks and metrics tailored for real-world deployment settings."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 47,
+ 398,
+ 299,
+ 517
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 47,
+ 398,
+ 299,
+ 517
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 398,
+ 299,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 398,
+ 299,
+ 517
+ ],
+ "type": "image",
+ "image_path": "7e5d6796694a8d1a706054c8a700440e2e26505bd752ab4e4efae519e5f05197.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 54,
+ 524,
+ 292,
+ 548
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 54,
+ 524,
+ 292,
+ 548
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 54,
+ 524,
+ 292,
+ 548
+ ],
+ "type": "text",
+ "content": "Fig. 10: The overview of evaluation and benchmarks in single LLM's deployment phase."
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 45,
+ 559,
+ 299,
+ 571
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 559,
+ 299,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 559,
+ 299,
+ 571
+ ],
+ "type": "text",
+ "content": "Robustness Evaluation. To systematically assess the relia-"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 46,
+ 607,
+ 299,
+ 744
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 48,
+ 580,
+ 297,
+ 604
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 580,
+ 297,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 580,
+ 297,
+ 604
+ ],
+ "type": "text",
+ "content": "TABLE 7: Summary of LLM robustness benchmarks at the deployment stage."
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 46,
+ 607,
+ 299,
+ 744
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 607,
+ 299,
+ 744
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 607,
+ 299,
+ 744
+ ],
+ "type": "table",
+ "html": "| Prior Work | Strategy | Description | Syntax Checker | Unit Tests |
| RTLCoder [7] | Keyword-based Generation, Mutation | Prompt LLM with keywords and existing code, followed by iterative mutation to get instruction-code pairs. | ✓ | X |
| OriGen [8] | Code-to-Code, Syntax Error Correction | Applies LLM-driven code-to-code pipeline on existing RTL code and filters them by compiler error feedback. | ✓ | X |
| BetterV [24] | Web Scraping & Cleaning, Alignment with C | Large-scale web-collected Verilog, cleaned and filtered to enforce coding standards; aligns C with Verilog. | ✓ | X |
| VeriGen [26] | Manually Collect Textbook and Open-Source Code | Mines real-world RTL from GitHub and textbooks, manually cleans and organizes them into a structured dataset. | ✓ | X |
| ChipGPT [27] | AST-based Synthesis | Converts Verilog ASTs into natural-language prompts and injects semantic error variants via EDA-tool feedback. | ✓ | X |
| VeriCoder (Our Work) | Feedback-Directed Refinement, Simulation, Unit Test Generation | Iteratively generate unit tests with a teacher LLM, check implementations via compiler and simulator, and refining designs and tests until each design passes. | ✓ | ✓ |
",
+ "bbox": [
+ 76,
+ 63,
+ 923,
+ 305
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "TABLE I: Comparison of Verilog fine-tuning dataset construction approaches.",
+ "bbox": [
+ 233,
+ 310,
+ 759,
+ 325
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "correctness| Prior Work | Strategy | Description | Syntax Checker | Unit Tests |
| RTLCoder [7] | Keyword-based Generation, Mutation | Prompt LLM with keywords and existing code, followed by iterative mutation to get instruction-code pairs. | ✓ | X |
| OriGen [8] | Code-to-Code, Syntax Error Correction | Applies LLM-driven code-to-code pipeline on existing RTL code and filters them by compiler error feedback. | ✓ | X |
| BetterV [24] | Web Scraping & Cleaning, Alignment with C | Large-scale web-collected Verilog, cleaned and filtered to enforce coding standards; aligns C with Verilog. | ✓ | X |
| VeriGen [26] | Manually Collect Textbook and Open-Source Code | Mines real-world RTL from GitHub and textbooks, manually cleans and organizes them into a structured dataset. | ✓ | X |
| ChipGPT [27] | AST-based Synthesis | Converts Verilog ASTs into natural-language prompts and injects semantic error variants via EDA-tool feedback. | ✓ | X |
| VeriCoder (Our Work) | Feedback-Directed Refinement, Simulation, Unit Test Generation | Iteratively generate unit tests with a teacher LLM, check implementations via compiler and simulator, and refining designs and tests until each design passes. | ✓ | ✓ |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.234,
+ 0.311,
+ 0.761,
+ 0.326
+ ],
+ "angle": 0,
+ "content": "TABLE I: Comparison of Verilog fine-tuning dataset construction approaches."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.074,
+ 0.353,
+ 0.492,
+ 0.549
+ ],
+ "angle": 0,
+ "content": "correctness| Prior Work | Strategy | Description | Syntax Checker | Unit Tests |
| RTLCoder [7] | Keyword-based Generation, Mutation | Prompt LLM with keywords and existing code, followed by iterative mutation to get instruction-code pairs. | ✓ | X |
| OriGen [8] | Code-to-Code, Syntax Error Correction | Applies LLM-driven code-to-code pipeline on existing RTL code and filters them by compiler error feedback. | ✓ | X |
| BetterV [24] | Web Scraping & Cleaning, Alignment with C | Large-scale web-collected Verilog, cleaned and filtered to enforce coding standards; aligns C with Verilog. | ✓ | X |
| VeriGen [26] | Manually Collect Textbook and Open-Source Code | Mines real-world RTL from GitHub and textbooks, manually cleans and organizes them into a structured dataset. | ✓ | X |
| ChipGPT [27] | AST-based Synthesis | Converts Verilog ASTs into natural-language prompts and injects semantic error variants via EDA-tool feedback. | ✓ | X |
| VeriCoder (Our Work) | Feedback-Directed Refinement, Simulation, Unit Test Generation | Iteratively generate unit tests with a teacher LLM, check implementations via compiler and simulator, and refining designs and tests until each design passes. | ✓ | ✓ |
+
+TABLE I: Comparison of Verilog fine-tuning dataset construction approaches.
+
+correctness| Prior Work | Strategy | Description | Syntax Checker | Unit Tests |
| RTLCoder [7] | Keyword-based Generation, Mutation | Prompt LLM with keywords and existing code, followed by iterative mutation to get instruction-code pairs. | ✓ | X |
| OriGen [8] | Code-to-Code, Syntax Error Correction | Applies LLM-driven code-to-code pipeline on existing RTL code and filters them by compiler error feedback. | ✓ | X |
| BetterV [24] | Web Scraping & Cleaning, Alignment with C | Large-scale web-collected Verilog, cleaned and filtered to enforce coding standards; aligns C with Verilog. | ✓ | X |
| VeriGen [26] | Manually Collect Textbook and Open-Source Code | Mines real-world RTL from GitHub and textbooks, manually cleans and organizes them into a structured dataset. | ✓ | X |
| ChipGPT [27] | AST-based Synthesis | Converts Verilog ASTs into natural-language prompts and injects semantic error variants via EDA-tool feedback. | ✓ | X |
| VeriCoder (Our Work) | Feedback-Directed Refinement, Simulation, Unit Test Generation | Iteratively generate unit tests with a teacher LLM, check implementations via compiler and simulator, and refining designs and tests until each design passes. | ✓ | ✓ |
",
+ "image_path": "9e6ce1035e375309829d6f9583b1aeaed11f21f05777f87cc4761e227314122f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 143,
+ 246,
+ 465,
+ 258
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 143,
+ 246,
+ 465,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 143,
+ 246,
+ 465,
+ 258
+ ],
+ "type": "text",
+ "content": "TABLE I: Comparison of Verilog fine-tuning dataset construction approaches."
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 45,
+ 279,
+ 301,
+ 434
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 279,
+ 301,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 279,
+ 301,
+ 434
+ ],
+ "type": "text",
+ "content": "correctness| Input image-text pair & Auxiliary sentences | Method | Semantic Rationale (SR of GT, Chimera and ChatGPT-4o) Reasoning Cause(RC of MDCA) | Impression Rationale (IR of GT and Chimera)/ Direct Cause(DC of MDCA) | Sentiment |
| RT @ MamiHerald : [Joanne Stiger]POS and Melissa Keller are the first same - sex couple to be married in Broward : | Ground Truth (GT) | Based on the image-text pair, the sentiment towards Joanne Stiger in the given text and image is positive, because the text highlights a historic and celebratory event of her marriage, and the image shows people smiling and appearing joyful, indicating happiness and approval. | The image emanates a joyous and celebratory atmosphere, highlighted by the smiling faces and relaxed postures of the individuals present. The presence of photographers and the gathering of people around the couple enhance the sense of an important and positive milestone being reached. Overall, the visual conveys happiness, love, and a significant moment being warmly embraced by all involved. | Positive |
| Chimera | The sentiment towards Joanne Stiger in the given text and image is positive, because the text highlights her as the first same-sex couple to be married in Broward, and the image shows her smiling and looking happy, which conveys a sense of pride and joy. | The image depicts two individuals who appear relaxed and happy, with one person appearing content and relaxed. The overall impression is one of happiness and contentment. The setting and body language suggest a sense of happiness, together with a pleasant and celebratory atmosphere. | Positive |
| Aesthetic caption: The body language of the two people shows that they are in a very different way. The background shows a room with a lot of people and the cameraman is in the background. | MDCA | Because the tweet simply reports on a factual event without expressing any emotional response towards Joanne Stiger. The use of the word "first" suggests that the author is excited about the news and is highlighting the fact that Stiger and Keller are the first same-sex couple to be married in Broward. Overall, the tweet does not provide enough information to determine a positive or negative sentiment towards JoAnn Stiger herself. | Because joanne stiger and Melissa keller are the first same - sex couple to be married in bronward. | Neutral |
| GPT-4o | Based on the image-text pair, the sentiment towards Joanne Stiger in the given text and image is Positive, because both the faces of Joanne Stiger and Melissa Keller, as well as the individuals around them, display expressions of happiness and joy, indicating a celebratory and momentous occasion. | | Positive |
| Primary Football Cup . [St Bede]' s v St Augustines friendly . . . both out but now for the friendly # sportsmanship | Ground Truth | Based on the image-text pair, the sentiment towards St Bede in the given text and image is neutral, because the text describes the event as a friendly match with a focus on sportsmanship, and the image shows a regular football game without any emotional or evaluative expressions towards St Bede. | The image depicts a friendly football match between St Bede's and St Augustine's on a lush green field under a partly cloudy sky. The overall scene conveys a sense of camaraderie and sportsmanship, aligning well with the neutral sentiment label. The focus appears to be on enjoying the game and fostering friendly competition in a relaxed atmosphere. | Neutral |
| Chimera | The sentiment towards St Bede in the given text and image is neutral, because the text simply states that St Bede is participating in a friendly match without expressing any positive or negative emotions, and the image shows a standard football match scene without any clear emotional cues. | The image of the St Bede vs. St Augustines football match conveys a sense of camaraderie and sportsmanship. The players are engaged in a moment of play, suggesting a positive and supportive atmosphere. The neutral sentiment label aligns with the overall feeling of cambered athleticism and teamwork, without evoking strong emotions. | Neutral |
| Aesthetic caption: The position of the camera captures the players in action, as they are engaged in a game of soccer. the position of the camera captures the players in action, and the position of the camera captures the players in action. | MDCA | Because the writer is promoting the team's friendly football match with St. Albers, which suggests that they are a fan of the team. The use of the hashtag #sportsmanship also suggests that the writer has a positive opinion of St. Bede. Overall, the language and tone of the tweet suggest that the sentiment towards St. Bede is positive. | It is a friendly football cup. | Positive |
| GPT-4o | Based on the image-text pair, the sentiment towards St Bede in the given text and image is Positive, because the text highlights a friendly and sportsmanlike attitude despite both teams being out of the competition, and the image shows players engaging in the match, fostering a positive environment. | | Positive |
| [Michael Oer]NEG is latest reminder that NFL pain lasts 24 / 7 / 365 | Ground Truth | Based on the image-text pair, the sentiment towards Michael Oer in the given text and image is negative, because the text highlights the constant and enduring pain associated with his NFL career, implying a prolonged and difficult experience. | The image of Michael Oer in his NFL uniform, combined with the sentiment label "negative", conveys a sense of enduring struggle and fatigue. His posture and facial expression suggest a feeling of exhaustion or frustration, which aligns with the text's implication that the pain and challenges of an NFL career are relentless and unforgiving. The overall impression is one of continuous physical and emotional burden. | Negative |
| Chimera | The sentiment towards Michael Oer in the given text and image is negative, because the text highlights that his recent reminder that NFL pain lasts 24/7/365, which is a serious and distressing reminder. | The image of Michael Oer exudes a sense of determination and focus, aligning with the negative sentiment label. His serious expression and the context of NFL pain underscore a feeling of exhaustion and exhaustion, which complements the idea of prolonged pain and the accompanying reminder that "no matter how much you love it," evoking a negative feeling. | Negative |
| MDCA | Because the tweet does not express any clear positive or negative opinion towards him. The tweet simply states that he is a reminder that NFL pain lasts for a specific time period. There is no indication of any emotional attachment or opinion towards Michael Oer in the tweet, and the language used is informative rather than emotional. Therefore, the sentiment towards Michael Oer in this tweet is neutral. | He is reminder that NFL pain lasts 24 / 7 / 365. | Neutral |
| GPT-4o | Based on the image-text pair, the sentiment towards Michael Oer in the given text and image is Negative, because the text highlights the unending pain associated with NFL careers, using Michael Oer as an example, and alludes to the physical and possibly emotional toll he endures. | | Negative |
",
+ "bbox": [
+ 96,
+ 54,
+ 901,
+ 694
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Fig. 6. Three examples showcasing the predictions generated by Chimera, MDCA, and GPT-4o are presented for analysis. During the evaluation process, GPT-4o exclusively produces the semantic rationale (SR). The input image-text pair and auxiliary sentences are utilized solely by Chimera. For MDCA, the reasoning cause (RC), direct cause (DC), and sentiment prediction are derived through direct inference.",
+ "bbox": [
+ 71,
+ 709,
+ 923,
+ 746
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "leads to the worst sentiment analysis performance across all datasets on IRG. This highlights the critical role of aesthetic captions in enhancing the model's understanding of image aesthetics, particularly in datasets like Twitter-2017 with balanced sentiment distributions. Specifically, attributes such as \"visual\" and \"vibrant\" positively contribute to sentiment analysis performance, whereas \"focus\" appears to significantly impair it. We speculate that since \"focus\" emphasizes specific image elements, potentially leads to an unbalanced interpretation of visual content. This localized emphasis can narrow the model's analytical scope, prioritizing details at",
+ "bbox": [
+ 71,
+ 770,
+ 491,
+ 931
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "the expense of broader context and compositional harmony. Consequently, the model may struggle to capture holistic aesthetic and emotional cues essential for accurate sentiment classification.",
+ "bbox": [
+ 503,
+ 770,
+ 921,
+ 829
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "5.4 Comparison with Large Language Models",
+ "text_level": 1,
+ "bbox": [
+ 504,
+ 849,
+ 856,
+ 864
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "We evaluate the performance of GPT-4o on the MASC task under a zero-shot setting. As shown in Table 7, GPT-4o achieves an accuracy of $46.87\\%$ and an F1 score of $47.47\\%$ , which is substantially lower than Chimera, which reports $81.61\\%$ accuracy and $77.98\\%$ F1 score. On the Twitter-2017",
+ "bbox": [
+ 503,
+ 869,
+ 923,
+ 941
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "header",
+ "text": "IEEE TRANSACTIONS ON AFFECTIVE COMPUTING, VOL. XX, NO. XX, XXXX 2025",
+ "bbox": [
+ 73,
+ 32,
+ 517,
+ 44
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "page_number",
+ "text": "14",
+ "bbox": [
+ 906,
+ 32,
+ 921,
+ 42
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "table",
+ "img_path": "images/f8457c3112501699ff84b3cfc411209dbad50c14d2e6528c484b1697dae5c507.jpg",
+ "table_caption": [
+ "TABLE 7 The experimental results $(\\%)$ of GPT-4o on the MASC task under a zero-shot setting are presented. The best-performing results highlighted in bold. The term \"dis\" refers to the percentage of samples where the sentiment polarity associated with a specific aspect cannot be discerned."
+ ],
+ "table_footnote": [],
+ "table_body": "| Input image-text pair & Auxiliary sentences | Method | Semantic Rationale (SR of GT, Chimera and ChatGPT-4o) Reasoning Cause(RC of MDCA) | Impression Rationale (IR of GT and Chimera)/ Direct Cause(DC of MDCA) | Sentiment |
| RT @ MamiHerald : [Joanne Stiger]POS and Melissa Keller are the first same - sex couple to be married in Broward : | Ground Truth (GT) | Based on the image-text pair, the sentiment towards Joanne Stiger in the given text and image is positive, because the text highlights a historic and celebratory event of her marriage, and the image shows people smiling and appearing joyful, indicating happiness and approval. | The image emanates a joyous and celebratory atmosphere, highlighted by the smiling faces and relaxed postures of the individuals present. The presence of photographers and the gathering of people around the couple enhance the sense of an important and positive milestone being reached. Overall, the visual conveys happiness, love, and a significant moment being warmly embraced by all involved. | Positive |
| Chimera | The sentiment towards Joanne Stiger in the given text and image is positive, because the text highlights her as the first same-sex couple to be married in Broward, and the image shows her smiling and looking happy, which conveys a sense of pride and joy. | The image depicts two individuals who appear relaxed and happy, with one person appearing content and relaxed. The overall impression is one of happiness and contentment. The setting and body language suggest a sense of happiness, together with a pleasant and celebratory atmosphere. | Positive |
| Aesthetic caption: The body language of the two people shows that they are in a very different way. The background shows a room with a lot of people and the cameraman is in the background. | MDCA | Because the tweet simply reports on a factual event without expressing any emotional response towards Joanne Stiger. The use of the word "first" suggests that the author is excited about the news and is highlighting the fact that Stiger and Keller are the first same-sex couple to be married in Broward. Overall, the tweet does not provide enough information to determine a positive or negative sentiment towards JoAnn Stiger herself. | Because joanne stiger and Melissa keller are the first same - sex couple to be married in bronward. | Neutral |
| GPT-4o | Based on the image-text pair, the sentiment towards Joanne Stiger in the given text and image is Positive, because both the faces of Joanne Stiger and Melissa Keller, as well as the individuals around them, display expressions of happiness and joy, indicating a celebratory and momentous occasion. | | Positive |
| Primary Football Cup . [St Bede]' s v St Augustines friendly . . . both out but now for the friendly # sportsmanship | Ground Truth | Based on the image-text pair, the sentiment towards St Bede in the given text and image is neutral, because the text describes the event as a friendly match with a focus on sportsmanship, and the image shows a regular football game without any emotional or evaluative expressions towards St Bede. | The image depicts a friendly football match between St Bede's and St Augustine's on a lush green field under a partly cloudy sky. The overall scene conveys a sense of camaraderie and sportsmanship, aligning well with the neutral sentiment label. The focus appears to be on enjoying the game and fostering friendly competition in a relaxed atmosphere. | Neutral |
| Chimera | The sentiment towards St Bede in the given text and image is neutral, because the text simply states that St Bede is participating in a friendly match without expressing any positive or negative emotions, and the image shows a standard football match scene without any clear emotional cues. | The image of the St Bede vs. St Augustines football match conveys a sense of camaraderie and sportsmanship. The players are engaged in a moment of play, suggesting a positive and supportive atmosphere. The neutral sentiment label aligns with the overall feeling of cambered athleticism and teamwork, without evoking strong emotions. | Neutral |
| Aesthetic caption: The position of the camera captures the players in action, as they are engaged in a game of soccer. the position of the camera captures the players in action, and the position of the camera captures the players in action. | MDCA | Because the writer is promoting the team's friendly football match with St. Albers, which suggests that they are a fan of the team. The use of the hashtag #sportsmanship also suggests that the writer has a positive opinion of St. Bede. Overall, the language and tone of the tweet suggest that the sentiment towards St. Bede is positive. | It is a friendly football cup. | Positive |
| GPT-4o | Based on the image-text pair, the sentiment towards St Bede in the given text and image is Positive, because the text highlights a friendly and sportsmanlike attitude despite both teams being out of the competition, and the image shows players engaging in the match, fostering a positive environment. | | Positive |
| [Michael Oer]NEG is latest reminder that NFL pain lasts 24 / 7 / 365 | Ground Truth | Based on the image-text pair, the sentiment towards Michael Oer in the given text and image is negative, because the text highlights the constant and enduring pain associated with his NFL career, implying a prolonged and difficult experience. | The image of Michael Oer in his NFL uniform, combined with the sentiment label "negative", conveys a sense of enduring struggle and fatigue. His posture and facial expression suggest a feeling of exhaustion or frustration, which aligns with the text's implication that the pain and challenges of an NFL career are relentless and unforgiving. The overall impression is one of continuous physical and emotional burden. | Negative |
| Chimera | The sentiment towards Michael Oer in the given text and image is negative, because the text highlights that his recent reminder that NFL pain lasts 24/7/365, which is a serious and distressing reminder. | The image of Michael Oer exudes a sense of determination and focus, aligning with the negative sentiment label. His serious expression and the context of NFL pain underscore a feeling of exhaustion and exhaustion, which complements the idea of prolonged pain and the accompanying reminder that "no matter how much you love it," evoking a negative feeling. | Negative |
| MDCA | Because the tweet does not express any clear positive or negative opinion towards him. The tweet simply states that he is a reminder that NFL pain lasts for a specific time period. There is no indication of any emotional attachment or opinion towards Michael Oer in the tweet, and the language used is informative rather than emotional. Therefore, the sentiment towards Michael Oer in this tweet is neutral. | He is reminder that NFL pain lasts 24 / 7 / 365. | Neutral |
| GPT-4o | Based on the image-text pair, the sentiment towards Michael Oer in the given text and image is Negative, because the text highlights the unending pain associated with NFL careers, using Michael Oer as an example, and alludes to the physical and possibly emotional toll he endures. | | Negative |
"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.073,
+ 0.71,
+ 0.924,
+ 0.747
+ ],
+ "angle": 0,
+ "content": "Fig. 6. Three examples showcasing the predictions generated by Chimera, MDCA, and GPT-4o are presented for analysis. During the evaluation process, GPT-4o exclusively produces the semantic rationale (SR). The input image-text pair and auxiliary sentences are utilized solely by Chimera. For MDCA, the reasoning cause (RC), direct cause (DC), and sentiment prediction are derived through direct inference."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.073,
+ 0.771,
+ 0.492,
+ 0.933
+ ],
+ "angle": 0,
+ "content": "leads to the worst sentiment analysis performance across all datasets on IRG. This highlights the critical role of aesthetic captions in enhancing the model's understanding of image aesthetics, particularly in datasets like Twitter-2017 with balanced sentiment distributions. Specifically, attributes such as \"visual\" and \"vibrant\" positively contribute to sentiment analysis performance, whereas \"focus\" appears to significantly impair it. We speculate that since \"focus\" emphasizes specific image elements, potentially leads to an unbalanced interpretation of visual content. This localized emphasis can narrow the model's analytical scope, prioritizing details at"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.504,
+ 0.771,
+ 0.923,
+ 0.83
+ ],
+ "angle": 0,
+ "content": "the expense of broader context and compositional harmony. Consequently, the model may struggle to capture holistic aesthetic and emotional cues essential for accurate sentiment classification."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.505,
+ 0.85,
+ 0.857,
+ 0.866
+ ],
+ "angle": 0,
+ "content": "5.4 Comparison with Large Language Models"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.504,
+ 0.87,
+ 0.924,
+ 0.943
+ ],
+ "angle": 0,
+ "content": "We evaluate the performance of GPT-4o on the MASC task under a zero-shot setting. As shown in Table 7, GPT-4o achieves an accuracy of \\(46.87\\%\\) and an F1 score of \\(47.47\\%\\), which is substantially lower than Chimera, which reports \\(81.61\\%\\) accuracy and \\(77.98\\%\\) F1 score. On the Twitter-2017"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.077,
+ 0.034,
+ 0.518,
+ 0.044
+ ],
+ "angle": 0,
+ "content": "IEEE TRANSACTIONS ON AFFECTIVE COMPUTING, VOL. XX, NO. XX, XXXX 2025"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.908,
+ 0.034,
+ 0.923,
+ 0.043
+ ],
+ "angle": 0,
+ "content": "15"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.078,
+ 0.056,
+ 0.488,
+ 0.124
+ ],
+ "angle": 0,
+ "content": "TABLE 7 The experimental results \\((\\%)\\) of GPT-4o on the MASC task under a zero-shot setting are presented. The best-performing results highlighted in bold. The term \"dis\" refers to the percentage of samples where the sentiment polarity associated with a specific aspect cannot be discerned."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.078,
+ 0.136,
+ 0.488,
+ 0.23
+ ],
+ "angle": 0,
+ "content": "| Input image-text pair & Auxiliary sentences | Method | Semantic Rationale (SR of GT, Chimera and ChatGPT-4o) Reasoning Cause(RC of MDCA) | Impression Rationale (IR of GT and Chimera)/ Direct Cause(DC of MDCA) | Sentiment |
| RT @ MamiHerald : [Joanne Stiger]POS and Melissa Keller are the first same - sex couple to be married in Broward : | Ground Truth (GT) | Based on the image-text pair, the sentiment towards Joanne Stiger in the given text and image is positive, because the text highlights a historic and celebratory event of her marriage, and the image shows people smiling and appearing joyful, indicating happiness and approval. | The image emanates a joyous and celebratory atmosphere, highlighted by the smiling faces and relaxed postures of the individuals present. The presence of photographers and the gathering of people around the couple enhance the sense of an important and positive milestone being reached. Overall, the visual conveys happiness, love, and a significant moment being warmly embraced by all involved. | Positive |
| Chimera | The sentiment towards Joanne Stiger in the given text and image is positive, because the text highlights her as the first same-sex couple to be married in Broward, and the image shows her smiling and looking happy, which conveys a sense of pride and joy. | The image depicts two individuals who appear relaxed and happy, with one person appearing content and relaxed. The overall impression is one of happiness and contentment. The setting and body language suggest a sense of happiness, together with a pleasant and celebratory atmosphere. | Positive |
| Aesthetic caption: The body language of the two people shows that they are in a very different way. The background shows a room with a lot of people and the cameraman is in the background. | MDCA | Because the tweet simply reports on a factual event without expressing any emotional response towards Joanne Stiger. The use of the word "first" suggests that the author is excited about the news and is highlighting the fact that Stiger and Keller are the first same-sex couple to be married in Broward. Overall, the tweet does not provide enough information to determine a positive or negative sentiment towards JoAnn Stiger herself. | Because joanne stiger and Melissa keller are the first same - sex couple to be married in bronward. | Neutral |
| GPT-4o | Based on the image-text pair, the sentiment towards Joanne Stiger in the given text and image is Positive, because both the faces of Joanne Stiger and Melissa Keller, as well as the individuals around them, display expressions of happiness and joy, indicating a celebratory and momentous occasion. | | Positive |
| Primary Football Cup . [St Bede]' s v St Augustines friendly . . . both out but now for the friendly # sportsmanship | Ground Truth | Based on the image-text pair, the sentiment towards St Bede in the given text and image is neutral, because the text describes the event as a friendly match with a focus on sportsmanship, and the image shows a regular football game without any emotional or evaluative expressions towards St Bede. | The image depicts a friendly football match between St Bede's and St Augustine's on a lush green field under a partly cloudy sky. The overall scene conveys a sense of camaraderie and sportsmanship, aligning well with the neutral sentiment label. The focus appears to be on enjoying the game and fostering friendly competition in a relaxed atmosphere. | Neutral |
| Chimera | The sentiment towards St Bede in the given text and image is neutral, because the text simply states that St Bede is participating in a friendly match without expressing any positive or negative emotions, and the image shows a standard football match scene without any clear emotional cues. | The image of the St Bede vs. St Augustines football match conveys a sense of camaraderie and sportsmanship. The players are engaged in a moment of play, suggesting a positive and supportive atmosphere. The neutral sentiment label aligns with the overall feeling of cambered athleticism and teamwork, without evoking strong emotions. | Neutral |
| Aesthetic caption: The position of the camera captures the players in action, as they are engaged in a game of soccer. the position of the camera captures the players in action, and the position of the camera captures the players in action. | MDCA | Because the writer is promoting the team's friendly football match with St. Albers, which suggests that they are a fan of the team. The use of the hashtag #sportsmanship also suggests that the writer has a positive opinion of St. Bede. Overall, the language and tone of the tweet suggest that the sentiment towards St. Bede is positive. | It is a friendly football cup. | Positive |
| GPT-4o | Based on the image-text pair, the sentiment towards St Bede in the given text and image is Positive, because the text highlights a friendly and sportsmanlike attitude despite both teams being out of the competition, and the image shows players engaging in the match, fostering a positive environment. | | Positive |
| [Michael Oer]NEG is latest reminder that NFL pain lasts 24 / 7 / 365 | Ground Truth | Based on the image-text pair, the sentiment towards Michael Oer in the given text and image is negative, because the text highlights the constant and enduring pain associated with his NFL career, implying a prolonged and difficult experience. | The image of Michael Oer in his NFL uniform, combined with the sentiment label "negative", conveys a sense of enduring struggle and fatigue. His posture and facial expression suggest a feeling of exhaustion or frustration, which aligns with the text's implication that the pain and challenges of an NFL career are relentless and unforgiving. The overall impression is one of continuous physical and emotional burden. | Negative |
| Chimera | The sentiment towards Michael Oer in the given text and image is negative, because the text highlights that his recent reminder that NFL pain lasts 24/7/365, which is a serious and distressing reminder. | The image of Michael Oer exudes a sense of determination and focus, aligning with the negative sentiment label. His serious expression and the context of NFL pain underscore a feeling of exhaustion and exhaustion, which complements the idea of prolonged pain and the accompanying reminder that "no matter how much you love it," evoking a negative feeling. | Negative |
| MDCA | Because the tweet does not express any clear positive or negative opinion towards him. The tweet simply states that he is a reminder that NFL pain lasts for a specific time period. There is no indication of any emotional attachment or opinion towards Michael Oer in the tweet, and the language used is informative rather than emotional. Therefore, the sentiment towards Michael Oer in this tweet is neutral. | He is reminder that NFL pain lasts 24 / 7 / 365. | Neutral |
| GPT-4o | Based on the image-text pair, the sentiment towards Michael Oer in the given text and image is Negative, because the text highlights the unending pain associated with NFL careers, using Michael Oer as an example, and alludes to the physical and possibly emotional toll he endures. | | Negative |
+
+Fig. 6. Three examples showcasing the predictions generated by Chimera, MDCA, and GPT-4o are presented for analysis. During the evaluation process, GPT-4o exclusively produces the semantic rationale (SR). The input image-text pair and auxiliary sentences are utilized solely by Chimera. For MDCA, the reasoning cause (RC), direct cause (DC), and sentiment prediction are derived through direct inference.
+
+leads to the worst sentiment analysis performance across all datasets on IRG. This highlights the critical role of aesthetic captions in enhancing the model's understanding of image aesthetics, particularly in datasets like Twitter-2017 with balanced sentiment distributions. Specifically, attributes such as "visual" and "vibrant" positively contribute to sentiment analysis performance, whereas "focus" appears to significantly impair it. We speculate that since "focus" emphasizes specific image elements, potentially leads to an unbalanced interpretation of visual content. This localized emphasis can narrow the model's analytical scope, prioritizing details at
+
+the expense of broader context and compositional harmony. Consequently, the model may struggle to capture holistic aesthetic and emotional cues essential for accurate sentiment classification.
+
+# 5.4 Comparison with Large Language Models
+
+We evaluate the performance of GPT-4o on the MASC task under a zero-shot setting. As shown in Table 7, GPT-4o achieves an accuracy of $46.87\%$ and an F1 score of $47.47\%$ , which is substantially lower than Chimera, which reports $81.61\%$ accuracy and $77.98\%$ F1 score. On the Twitter-2017
+
+TABLE 7 The experimental results $(\%)$ of GPT-4o on the MASC task under a zero-shot setting are presented. The best-performing results highlighted in bold. The term "dis" refers to the percentage of samples where the sentiment polarity associated with a specific aspect cannot be discerned.
+
+| Input image-text pair & Auxiliary sentences | Method | Semantic Rationale (SR of GT, Chimera and ChatGPT-4o) Reasoning Cause(RC of MDCA) | Impression Rationale (IR of GT and Chimera)/ Direct Cause(DC of MDCA) | Sentiment |
| RT @ MamiHerald : [Joanne Stiger]POS and Melissa Keller are the first same - sex couple to be married in Broward : | Ground Truth (GT) | Based on the image-text pair, the sentiment towards Joanne Stiger in the given text and image is positive, because the text highlights a historic and celebratory event of her marriage, and the image shows people smiling and appearing joyful, indicating happiness and approval. | The image emanates a joyous and celebratory atmosphere, highlighted by the smiling faces and relaxed postures of the individuals present. The presence of photographers and the gathering of people around the couple enhance the sense of an important and positive milestone being reached. Overall, the visual conveys happiness, love, and a significant moment being warmly embraced by all involved. | Positive |
| Chimera | The sentiment towards Joanne Stiger in the given text and image is positive, because the text highlights her as the first same-sex couple to be married in Broward, and the image shows her smiling and looking happy, which conveys a sense of pride and joy. | The image depicts two individuals who appear relaxed and happy, with one person appearing content and relaxed. The overall impression is one of happiness and contentment. The setting and body language suggest a sense of happiness, together with a pleasant and celebratory atmosphere. | Positive |
| Aesthetic caption: The body language of the two people shows that they are in a very different way. The background shows a room with a lot of people and the cameraman is in the background. | MDCA | Because the tweet simply reports on a factual event without expressing any emotional response towards Joanne Stiger. The use of the word "first" suggests that the author is excited about the news and is highlighting the fact that Stiger and Keller are the first same-sex couple to be married in Broward. Overall, the tweet does not provide enough information to determine a positive or negative sentiment towards JoAnn Stiger herself. | Because joanne stiger and Melissa keller are the first same - sex couple to be married in bronward. | Neutral |
| GPT-4o | Based on the image-text pair, the sentiment towards Joanne Stiger in the given text and image is Positive, because both the faces of Joanne Stiger and Melissa Keller, as well as the individuals around them, display expressions of happiness and joy, indicating a celebratory and momentous occasion. | | Positive |
| Primary Football Cup . [St Bede]' s v St Augustines friendly . . . both out but now for the friendly # sportsmanship | Ground Truth | Based on the image-text pair, the sentiment towards St Bede in the given text and image is neutral, because the text describes the event as a friendly match with a focus on sportsmanship, and the image shows a regular football game without any emotional or evaluative expressions towards St Bede. | The image depicts a friendly football match between St Bede's and St Augustine's on a lush green field under a partly cloudy sky. The overall scene conveys a sense of camaraderie and sportsmanship, aligning well with the neutral sentiment label. The focus appears to be on enjoying the game and fostering friendly competition in a relaxed atmosphere. | Neutral |
| Chimera | The sentiment towards St Bede in the given text and image is neutral, because the text simply states that St Bede is participating in a friendly match without expressing any positive or negative emotions, and the image shows a standard football match scene without any clear emotional cues. | The image of the St Bede vs. St Augustines football match conveys a sense of camaraderie and sportsmanship. The players are engaged in a moment of play, suggesting a positive and supportive atmosphere. The neutral sentiment label aligns with the overall feeling of cambered athleticism and teamwork, without evoking strong emotions. | Neutral |
| Aesthetic caption: The position of the camera captures the players in action, as they are engaged in a game of soccer. the position of the camera captures the players in action, and the position of the camera captures the players in action. | MDCA | Because the writer is promoting the team's friendly football match with St. Albers, which suggests that they are a fan of the team. The use of the hashtag #sportsmanship also suggests that the writer has a positive opinion of St. Bede. Overall, the language and tone of the tweet suggest that the sentiment towards St. Bede is positive. | It is a friendly football cup. | Positive |
| GPT-4o | Based on the image-text pair, the sentiment towards St Bede in the given text and image is Positive, because the text highlights a friendly and sportsmanlike attitude despite both teams being out of the competition, and the image shows players engaging in the match, fostering a positive environment. | | Positive |
| [Michael Oer]NEG is latest reminder that NFL pain lasts 24 / 7 / 365 | Ground Truth | Based on the image-text pair, the sentiment towards Michael Oer in the given text and image is negative, because the text highlights the constant and enduring pain associated with his NFL career, implying a prolonged and difficult experience. | The image of Michael Oer in his NFL uniform, combined with the sentiment label "negative", conveys a sense of enduring struggle and fatigue. His posture and facial expression suggest a feeling of exhaustion or frustration, which aligns with the text's implication that the pain and challenges of an NFL career are relentless and unforgiving. The overall impression is one of continuous physical and emotional burden. | Negative |
| Chimera | The sentiment towards Michael Oer in the given text and image is negative, because the text highlights that his recent reminder that NFL pain lasts 24/7/365, which is a serious and distressing reminder. | The image of Michael Oer exudes a sense of determination and focus, aligning with the negative sentiment label. His serious expression and the context of NFL pain underscore a feeling of exhaustion and exhaustion, which complements the idea of prolonged pain and the accompanying reminder that "no matter how much you love it," evoking a negative feeling. | Negative |
| MDCA | Because the tweet does not express any clear positive or negative opinion towards him. The tweet simply states that he is a reminder that NFL pain lasts for a specific time period. There is no indication of any emotional attachment or opinion towards Michael Oer in the tweet, and the language used is informative rather than emotional. Therefore, the sentiment towards Michael Oer in this tweet is neutral. | He is reminder that NFL pain lasts 24 / 7 / 365. | Neutral |
| GPT-4o | Based on the image-text pair, the sentiment towards Michael Oer in the given text and image is Negative, because the text highlights the unending pain associated with NFL careers, using Michael Oer as an example, and alludes to the physical and possibly emotional toll he endures. | | Negative |
",
+ "image_path": "76169821fecec912e1e9076d4013b612fbcbf9e394366e9d11fb88f543258d70.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 44,
+ 562,
+ 565,
+ 591
+ ],
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 562,
+ 565,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 562,
+ 565,
+ 591
+ ],
+ "type": "text",
+ "content": "Fig. 6. Three examples showcasing the predictions generated by Chimera, MDCA, and GPT-4o are presented for analysis. During the evaluation process, GPT-4o exclusively produces the semantic rationale (SR). The input image-text pair and auxiliary sentences are utilized solely by Chimera. For MDCA, the reasoning cause (RC), direct cause (DC), and sentiment prediction are derived through direct inference."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 44,
+ 610,
+ 301,
+ 738
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 44,
+ 610,
+ 301,
+ 738
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 44,
+ 610,
+ 301,
+ 738
+ ],
+ "type": "text",
+ "content": "leads to the worst sentiment analysis performance across all datasets on IRG. This highlights the critical role of aesthetic captions in enhancing the model's understanding of image aesthetics, particularly in datasets like Twitter-2017 with balanced sentiment distributions. Specifically, attributes such as \"visual\" and \"vibrant\" positively contribute to sentiment analysis performance, whereas \"focus\" appears to significantly impair it. We speculate that since \"focus\" emphasizes specific image elements, potentially leads to an unbalanced interpretation of visual content. This localized emphasis can narrow the model's analytical scope, prioritizing details at"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 308,
+ 610,
+ 564,
+ 657
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 610,
+ 564,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 610,
+ 564,
+ 657
+ ],
+ "type": "text",
+ "content": "the expense of broader context and compositional harmony. Consequently, the model may struggle to capture holistic aesthetic and emotional cues essential for accurate sentiment classification."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 309,
+ 673,
+ 524,
+ 685
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 309,
+ 673,
+ 524,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 673,
+ 524,
+ 685
+ ],
+ "type": "text",
+ "content": "5.4 Comparison with Large Language Models"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 308,
+ 689,
+ 565,
+ 746
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 689,
+ 565,
+ 746
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 689,
+ 565,
+ 746
+ ],
+ "type": "text",
+ "content": "We evaluate the performance of GPT-4o on the MASC task under a zero-shot setting. As shown in Table 7, GPT-4o achieves an accuracy of "
+ },
+ {
+ "bbox": [
+ 308,
+ 689,
+ 565,
+ 746
+ ],
+ "type": "inline_equation",
+ "content": "46.87\\%"
+ },
+ {
+ "bbox": [
+ 308,
+ 689,
+ 565,
+ 746
+ ],
+ "type": "text",
+ "content": " and an F1 score of "
+ },
+ {
+ "bbox": [
+ 308,
+ 689,
+ 565,
+ 746
+ ],
+ "type": "inline_equation",
+ "content": "47.47\\%"
+ },
+ {
+ "bbox": [
+ 308,
+ 689,
+ 565,
+ 746
+ ],
+ "type": "text",
+ "content": ", which is substantially lower than Chimera, which reports "
+ },
+ {
+ "bbox": [
+ 308,
+ 689,
+ 565,
+ 746
+ ],
+ "type": "inline_equation",
+ "content": "81.61\\%"
+ },
+ {
+ "bbox": [
+ 308,
+ 689,
+ 565,
+ 746
+ ],
+ "type": "text",
+ "content": " accuracy and "
+ },
+ {
+ "bbox": [
+ 308,
+ 689,
+ 565,
+ 746
+ ],
+ "type": "inline_equation",
+ "content": "77.98\\%"
+ },
+ {
+ "bbox": [
+ 308,
+ 689,
+ 565,
+ 746
+ ],
+ "type": "text",
+ "content": " F1 score. On the Twitter-2017"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 45,
+ 26,
+ 317,
+ 35
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 26,
+ 317,
+ 35
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 26,
+ 317,
+ 35
+ ],
+ "type": "text",
+ "content": "IEEE TRANSACTIONS ON AFFECTIVE COMPUTING, VOL. XX, NO. XX, XXXX 2025"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 555,
+ 26,
+ 564,
+ 34
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 555,
+ 26,
+ 564,
+ 34
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 555,
+ 26,
+ 564,
+ 34
+ ],
+ "type": "text",
+ "content": "14"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 13
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 47,
+ 107,
+ 298,
+ 182
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 47,
+ 44,
+ 298,
+ 98
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 44,
+ 298,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 44,
+ 298,
+ 98
+ ],
+ "type": "text",
+ "content": "TABLE 7 The experimental results "
+ },
+ {
+ "bbox": [
+ 47,
+ 44,
+ 298,
+ 98
+ ],
+ "type": "inline_equation",
+ "content": "(\\%)"
+ },
+ {
+ "bbox": [
+ 47,
+ 44,
+ 298,
+ 98
+ ],
+ "type": "text",
+ "content": " of GPT-4o on the MASC task under a zero-shot setting are presented. The best-performing results highlighted in bold. The term \"dis\" refers to the percentage of samples where the sentiment polarity associated with a specific aspect cannot be discerned."
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 47,
+ 107,
+ 298,
+ 182
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 107,
+ 298,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 107,
+ 298,
+ 182
+ ],
+ "type": "table",
+ "html": "