acceptodds
Under review as a conference paper at ICLR 2027

Revealing Hidden Intent for Defending Text-to-Image Models against Adversarial Attacks

Abstract

Adversarial attacks against Text-to-Image (T2I) models can induce unsafe imagery while concealing underlying harmful intent through obfuscated perturbations. Existing defenses primarily mitigate such threats either by suppressing predefined unsafe patterns via interventions in the T2I generation process or by evaluating manipulated prompts with auxiliary signals, such as latent concept similarity and reconstruction discrepancy. However, their effectiveness can deteriorate when harmful content falls outside the targeted space or when such cues provide insufficient evidence of semantics rendered opaque by adversarial perturbations, underscoring the need to reveal the semantics these manipulations obscure. In this paper, we introduce EEGIS, a T2I defense framework that surfaces the latent intent embedded in adversarial prompts as semantically coherent descriptions for reliable and interpretable safety assessment. EEGIS reconstructs the concealed semantics of lexically corrupted expressions by inferring the visual content evoked by each perturbed token from the diffusion noise during early denoising. Building on the recovered semantics, it further grounds implicit implications into visually consistent explicit alternatives, exposing harmful intent that remains concealed beyond surface-level lexical distortions. Extensive experiments on NSFW-200 and T2I-RiskyPrompt across diverse attack methods demonstrate that EEGIS effectively captures adversarial intent while maintaining low false positive rates.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Text-to-Image Models, AI Safety, Adversarial Attacks, Adversarial Defense, Robustness

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_u7o5OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes EEGIS, a defense framework that protects Text-to-Image models by uncovering hidden adversarial intent behind manipulated prompts. Instead of relying on rigid keyword filters or concept suppression, EEGIS reconstructs obscured semantics from early diffusion noise and rewrites deceptive phrases into explicit descriptions for safety checks. Experiments on NSFW-200 and T2I-RiskyPrompt show it effectively catches diverse adversarial attacks with low false positive rates.

Critical strengths
  1. The proposed framework significantly enhances the interpretability of adversarial prompt defenses in T2I models by converting perturbed inputs into explicit visual descriptions.

  2. By leveraging visually guided semantic restoration rather than static concept lists, the algorithm demonstrates strong adaptability to unseen malicious concepts.

Critical weaknesses
  1. The paper lacks ablation experiments on the key hyperparameter kk, and the authors do not analyze how ASR and FPR vary with different values of kk; a clear quantitative guideline for setting kk across different T2I models should be provided.
  2. The Explicit Visual Grounding module plays a critical role in the ablation study, but implementing it via GPT-5.2 incurs high computational costs that hinder practical deployment, and the authors do not further explore the feasibility of using smaller models.
  3. The authors do not evaluate the framework against adaptive adversaries, leaving the worst-case performance of both proposed modules unverified.
Questions for authors
  1. Using GPT-5.2 for Explicit Visual Grounding is expensive and slow for real-world use. Can smaller, open-source models work as well here, and how robust is the defense if an attacker knows how your system works (adaptive attack)?
  2. How do ASR and FPR change when varying the hyperparameter kk? Could you provide experimental results for different values of kk and explain how to choose this parameter for other T2I models?
3 comments
AuthorsOpenReview ↗

Response to comments on the contextual rank threshold sensitivity

We sincerely appreciate the reviewer's valuable feedback and the time dedicated to reviewing our work. We have carefully considered the comments and hope that our additional experiments and explanations address the concerns raised.

In summary, we provide additional analyses of rank threshold sensitivity, evaluate EEGIS with smaller open-source LLMs and assess its API cost, and clarify the scope of our study regarding adaptive adversaries.

We provide detailed responses to each comment below.

W1: The paper lacks ablation experiments on the key hyperparameter k, and the authors do not analyze how ASR and FPR vary with different values of k; a clear quantitative guideline for setting k across different T2I models should be provided.

Q2: How do ASR and FPR change when varying the hyperparameter k? Could you provide experimental results for different values of k and explain how to choose this parameter for other T2I models?

We would like to clarify that EEGIS does not utilize a hyperparameter denoted by kk. We understand the kk mentioned in your comment to refer to the contextual rank threshold τ\tau introduced in Section 3.1, which determines the tokens selected for semantic verbalization.

Following your suggestion, we conducted a sensitivity analysis of τ\tau to examine its effects on ASR and FPR.

Attack Success Rates (ASR, ↓)

SneakyPromptMMA-DiffusionRing-A-BellP4DPGJ
τ=100\tau = 1002.865.520.551.744.42
τ=1000\tau = 10002.294.420.001.163.31
τ=5000\tau = 50002.295.520.001.164.42

 

False Positive Rates (FPR, ↓)

LAION-COCODiffusionDB
τ=100\tau = 1002.008.75
τ=1000\tau = 10001.506.50
τ=5000\tau = 50000.756.50

As shown in the tables, EEGIS maintains low ASR across all five attack methods at the evaluated thresholds, while FPR generally decreases as τ\tau increases. These results are consistent with the complementary roles of semantic verbalization and explicit visual grounding, which can reveal harmful implications even when no suspicious tokens are detected. Overall, EEGIS maintains effective defense performance across different threshold settings, providing a favorable balance between ASR and FPR. We have added these results and further discussion in Section 4.5 (L498-525).

We would like to clarify that the contextual rank threshold is applied to token ranks computed by the masked LM, rather than directly depending on the T2I backbone. Therefore, the same threshold can serve as an initial setting across different T2I models when using the same masked LM. For other settings, the threshold can be adjusted based on the ASR–FPR trade-off on a validation set, depending on the desired balance between defense effectiveness and false positives.

AuthorsOpenReview ↗

Response to comments on the feasibility of EEGIS with open-source LLMs

W2: The Explicit Visual Grounding module plays a critical role in the ablation study, but implementing it via GPT-5.2 incurs high computational costs that hinder practical deployment, and the authors do not further explore the feasibility of using smaller models. Q1: Using GPT-5.2 for Explicit Visual Grounding is expensive and slow for real-world use. Can smaller, open-source models work as well here?

Following your suggestion, we further evaluated EEGIS by replacing GPT-5.2 with two smaller open-source LLMs, Qwen-3.5-9B and GPT-OSS-20B, as shown in the tables below.

Attack Success Rates (ASR, ↓)

MethodSneakyPromptMMA-DiffusionRing-A-BellP4DPGJAVG. (↓)
ESD14.2927.6254.1447.5411.6031.04
SLD-Medium48.0074.5999.4582.9042.5469.50
SLD-Strong38.2964.6499.4580.5836.4663.88
SLD-Max26.8661.3391.7175.9424.3156.03
RECE17.7155.8039.7865.5121.5540.07
SAFREE14.8648.0797.7968.7018.7849.64
MACE12.5710.5013.8115.9414.3613.44
Latent Guard18.2930.9445.3058.844.4231.56
GuardT2I9.714.420.552.616.084.67
EEGIS (Qwen-3.5-9B)4.578.840.001.747.184.47
EEGIS (GPT-OSS-20B)4.009.940.001.455.524.18
EEGIS (GPT-5.2)2.294.420.001.163.312.24

 

False Positive Rates (FPR, ↓)

MethodLAION-COCODiffusionDB
ESD0.754.75
SLD-Medium2.255.75
SLD-Strong1.257.00
SLD-Max2.756.50
RECE3.003.75
SAFREE1.003.75
MACE3.258.50
Latent Guard14.0018.25
GuardT2I7.5053.00
EEGIS (Qwen-3.5-9B)0.003.25
EEGIS (GPT-OSS-20)0.005.00
EEGIS (GPT-5.2)1.506.50

Both open-source variants achieve relatively lower average ASR than GuardT2I, the strongest baseline, despite some performance degradation relative to GPT-5.2. Furthermore, they maintain low FPRs on both LAION-COCO and DiffusionDB, indicating that competitive defense performance can be achieved without substantially increasing false positives. These results suggest that EEGIS remains effective with smaller open-source LLMs, supporting its applicability beyond a specific proprietary model. We have added experimental results and further discussion in Appendix C (L1005-1013).

In addition, to address the computational cost concern, we measured the GPT-5.2 API cost incurred by EEGIS across adversarial prompts, as shown below.

API Cost of EEGIS with GPT-5.2 (USD per Prompt, ↓)

SneakyPromptMMA-DiffusionRing-A-BellP4DPGJAVG.
0.003380.003080.000340.002240.003450.00231

EEGIS incurs an average API cost of $0.00231 per prompt. The cost varies across attack methods, partly because prompts identified by early NSFW keyword filtering do not require subsequent LLM calls. This analysis quantifies the additional API cost associated with GPT-5.2, complementing the runtime analysis reported in our original evaluation. We have added these results and discussions in Appendix C (L959-962).

AuthorsOpenReview ↗

Response to comments on robustness against adaptive adversaries

W3: The authors do not evaluate the framework against adaptive adversaries, leaving the worst-case performance of both proposed modules unverified.

Q1: how robust is the defense if an attacker knows how your system works (adaptive attack)?

Our study focuses on revealing concealed harmful intent across diverse adversarial attacks, rather than specifically defending against adaptive adversaries. Accordingly, we evaluate EEGIS against five established T2I attack methods with distinct mechanisms (Table 1).

We acknowledge that defending against adaptive adversaries is an important research direction. However, evaluating worst-case robustness against such attacks would require a dedicated defense-aware threat model, attacks specifically designed to circumvent EEGIS, and an appropriate evaluation protocol, which fall outside the primary scope of our study. We recognize this as a valuable direction for future research.

Reviewer_GhmTOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EEGIS, a defense for text-to-image models that decides whether a prompt is unsafe before image generation. It flags suspicious tokens with a masked language model, recovers the intended word of each flagged token from early-denoising diffusion features, and then has an LLM rewrite euphemistic expressions into explicit candidates. An LLM judge classifies the prompt from the recovered prompt and the candidates. The paper claims that this reveals hidden intent and gives reliable and interpretable safety assessment under diverse adversarial attacks with low false positive rates.

Critical strengths
  1. The defense outputs recovered and explicit text instead of a suspicion score, which addresses the gap the paper identifies in Latent Guard and GuardT2I. Table 1 shows the lowest or tied-lowest ASR on all five attacks, so the gain is not confined to one attack type.
  2. The evaluation reports false positive rate on two benign sets, runtime, and a human check of the LLM judge in the appendix. These reports let readers see the cost side of the defense and not only the ASR.
Critical weaknesses
  1. In Section 4.3, the paper claims that restored and candidate prompts reveal hidden intent, but Table 3 shows restored prompts differ from adversarial prompts by 0.003 in BERTScore and 0.010 in SigLIP, and candidates have lower SigLIP. The authors should measure fidelity to the original intent with human ratings.
  2. In Section 3.1, prompts with no token above the fixed masked-LM rank threshold skip semantic verbalization, so for lexically plausible attacks such as PGJ the decision rests on the LLM stages (Table 4). The first stage is a perplexity-style signal that is not compared, so the authors should add a perplexity-filter baseline and threshold sensitivity.
  3. In Sections 3.2 and 3.3, no baseline trains a classifier on early-denoising features, and the final assessment also uses the NSFW keyword list inherited from GuardT2I. The need for text recovery and explicit rewriting is not shown, so the authors should add a feature-probe baseline and ablate the keyword list.
  4. In Section 3.2 and Appendix B, the restoration head predicts each flagged token independently by arg max and is trained with one masked word per clean prompt, while attacks such as P4D corrupt many tokens at once, and the restored prompt in Figure 3b is incoherent. The paper reports no token-level recovery accuracy, so the authors should evaluate recovery accuracy and train with multiple corrupted tokens.
  5. In Section 4.1, the evaluation uses one target model, its safety checker as ground truth, and one LLM judge whose identity is not stated, with no confidence intervals on gaps of a few prompts. The authors should add a second target model and classifier.
Questions for authors
  1. For each attack, what fraction of adversarial prompts contains at least one token above the threshold, and how does ASR change when the threshold varies?
  2. Which LLM serves as the final judge, and what is the ASR when the NSFW keyword list inherited from GuardT2I is removed?
4 comments
AuthorsOpenReview ↗

Response to comments on fidelity to original harmful intent and human rating

We sincerely appreciate the reviewer's time and effort in reviewing our work and providing constructive feedback. We have carefully considered the comments and hope that our additional experiments and clarifications address the concerns raised.

In summary, we conducted a human evaluation of semantic and harmful intent fidelity, introduced alternative detection baselines, analyzed contextual rank threshold sensitivity, and evaluated the impact of removing NSFW keyword filtering. These additional analyses further clarify the roles of Semantic Verbalization and Explicit Visual Grounding in EEGIS.

We provide detailed responses to each comment below.

W1: In Section 4.3, the paper claims that restored and candidate prompts reveal hidden intent, but Table 3 shows restored prompts differ from adversarial prompts by 0.003 in BERTScore and 0.010 in SigLIP, and candidates have lower SigLIP. The authors should measure fidelity to the original intent with human ratings.

We would like to clarify that Table 3 aims to assess how well the semantic and visual content of the original harmful prompts is preserved through semantic verbalization and explicit visual grounding, rather than demonstrate improvements in similarity over adversarial prompts. We thank the reviewer for pointing out this potential confusion. We removed the upward arrows from BERTScore, SigLIP, and Explicit and clarified the interpretation in Section 4.3 (L456-459).

In addition, following your suggestion, we further conducted a human evaluation to assess fidelity to the original harmful intent. We randomly sampled 30 prompt pairs processed by EEGIS from each attack method, except for Ring-A-Bell. For Ring-A-Bell, only 15 cases yield candidate prompts through explicit visual grounding, so all 15 are included in the evaluation. Four annotators who were not involved in this research independently compared images generated from restored and candidate prompts with those from the corresponding harmful prompts.

Using a five point scale, they evaluated (1) semantic fidelity, measuring preservation of the original scene, actions, and interactions, and (2) harmful intent fidelity, measuring preservation of the original harmful actions and intent. (Instructions on the five-point rating criteria for evaluation are provided in Appendix D (L1080-1100))

Human evaluation of semantic and harmful intent fidelity (average scores; higher is better)

EvaluationRestored ImageCandidate Image
Semantic fidelity4.18 ± 0.344.20 ± 0.34
Harmful intent fidelity4.36 ± 0.164.46 ± 0.14

As shown in the table, restored and candidate images achieve nearly identical semantic fidelity scores of 4.18 and 4.20, suggesting that the original scene and interactions are largely preserved. Meanwhile, harmful intent fidelity increases slightly from 4.36 to 4.46, suggesting that candidate images faithfully reflect the specific harmful actions and intent depicted in the original images. This suggests that explicit visual grounding preserves the underlying scene while making concealed harmful content more explicit. The human ratings further complement the similarity metrics in Table 3 by further confirming that the original scene and harmful intent are largely preserved in the generated images. Example image pairs for human rating and further discussion are provided in Appendix D (L1119-1167).

AuthorsOpenReview ↗

Response to comments on alternative detection baselines and rank threshold sensitivity

W2: In Section 3.1, prompts with no token above the fixed masked-LM rank threshold skip semantic verbalization, so for lexically plausible attacks such as PGJ the decision rests on the LLM stages (Table 4). The first stage is a perplexity-style signal that is not compared, so the authors should add a perplexity-filter baseline and threshold sensitivity.

W3: In Sections 3.2 and 3.3, no baseline trains a classifier on early-denoising features (…)

Following your suggestion, we conducted additional experiments comparing EEGIS with linguistic plausibility-based detection baselines and a classifier trained on early-denoising features, under an identical FPR to EEGIS. The results are as shown in the tables below.

Comparison with alternative detection baselines (ASR, ↓).

MethodSneakyPromptMMA-DiffusionRing-A-BellP4DPGJAVG. (↓)
Max Token62.8659.6761.3356.3552.7558.59
Contextual Rank58.2924.3115.4760.7734.2038.61
Perplexity69.7116.571.6660.7752.4640.23
Denoising50.2948.0739.2329.5748.0743.05
EEGIS2.294.420.001.163.312.24

Max Token detects suspicious prompts using the lowest contextual token likelihood within a prompt, while Contextual Rank uses the highest token rank predicted by a masked LM. Perplexity assesses the overall linguistic plausibility of a prompt, whereas Denoising directly classifies early-denoising representations using a trained binary classifier. For a fair comparison, the Denoising classifier is trained on CoPro and UnsafeBench using the same dataset as the restoration head in EEGIS.

Linguistic baselines exhibit varying performance across attack methods. In particular, Perplexity achieves low ASR on Ring-A-Bell, suggesting that linguistic plausibility can provide useful detection signals for certain adversarial perturbations. Similarly, the Denoising baseline shows relatively effective detection on P4D, indicating that early-denoising representations contain information relevant to identifying unsafe prompts. However, neither linguistic plausibility nor direct classification of early-denoising features provides consistently effective detection across attacks. In contrast, EEGIS leverages these signals for semantic recovery and further makes implicit harmful implications explicit, achieving consistently low ASR across diverse adversarial strategies. We have added these experimental results and further discussion in Appendix C (L1015-1055).

Q1: how does ASR change when the threshold varies?

To address this question, we additionally conducted a sensitivity analysis of the contextual rank threshold (τ\tau), which determines which tokens undergo semantic verbalization. As shown in the tables below.

Attack Success Rates (ASR, ↓)

SneakyPromptMMA-DiffusionRing-A-BellP4DPGJ
τ=100\tau = 1002.865.520.551.744.42
τ=1000\tau = 10002.294.420.001.163.31
τ=5000\tau = 50002.295.520.001.164.42

 

False Positive Rates (FPR, ↓)

LAION-COCODiffusionDB
τ=100\tau = 1002.008.75
τ=1000\tau = 10001.506.50
τ=5000\tau = 50000.756.50

As shown in the tables, EEGIS maintains low ASR across all five attack methods at the evaluated thresholds, while FPR generally decreases as τ\tau increases. These results are consistent with the complementary roles of semantic verbalization and explicit visual grounding, which can reveal harmful implications even when no suspicious tokens are detected. Overall, EEGIS maintains effective defense performance across different threshold settings, providing a favorable balance between ASR and FPR. We have added these results and further discussion in Section 4.5 (L498-525).

AuthorsOpenReview ↗

Response to comments on NSFW keyword filtering and token restoration accuracy

W3: In Sections 3.2 and 3.3, no (…) the final assessment also uses the NSFW keyword list inherited from GuardT2I. The need for text recovery and explicit rewriting is not shown, so the authors should add a feature-probe baseline and ablate the keyword list.

Q2: What is the ASR when the NSFW keyword list inherited from GuardT2I is removed?

To address this concern, we evaluated EEGIS and GuardT2I (the strongest baseline) with predefined NSFW keyword filtering disabled, as shown below (†\dagger indicates that predefined NSFW keyword filtering is not applied.).

Attack Success Rates (ASR, ↓)

MethodSneakyPromptMMA-DiffusionRing-A-BellP4DPGJAVG. (↓)
GuardT2I$^{\dagger}$13.717.736.637.837.738.73
GuardT2I9.714.420.552.616.084.67
EEGIS†^{\dagger}2.297.180.552.036.083.63
EEGIS2.294.420.001.163.312.24

Although removing NSFW keyword filtering increases ASR for both methods, EEGIS exhibits a smaller performance degradation. Notably, even without keyword filtering, EEGIS achieves a lower average ASR than GuardT2I with keyword filtering enabled, while maintaining the best or competitive performance across attack methods compared with other defense methods. These results indicate that the effectiveness of EEGIS is not solely dependent on predefined NSFW keyword filtering to identify concealed harmful intent. We have added these experimental results and further discussion in Appendix C (L998-1004).

W4: In Section 3.2 and Appendix B, the restoration head predicts each flagged token independently by arg max and is trained with one masked word per clean prompt, while attacks such as P4D corrupt many tokens at once, and the restored prompt in Figure 3b is incoherent. The paper reports no token-level recovery accuracy, so the authors should evaluate recovery accuracy and train with multiple corrupted tokens.

We would like to note that attacks such as P4D introduce multiple optimized tokens that do not have a clear one-to-one correspondence with the original tokens, making it difficult to define token-level restoration targets for supervised training. Moreover, training directly on such perturbations could introduce dependencies on specific attack mechanisms. Therefore, we train the restoration head on clean prompts to learn the relationship between textual tokens and visual representations without relying on attack-specific perturbations. We have added these clarifications in Appendix B (L861-889).

Importantly, semantic verbalization aims to recover obscured visual semantics rather than reconstruct the exact original tokens. As illustrated in Figure 3b, P4D substantially alters the prompt through token substitutions and insertions, making exact reconstruction particularly challenging. Although the restored prompt is not fully coherent, expressions such as "touching," "physical," and "sexual" remain consistent with the original scene, helping reveal the underlying harmful implications.

Following your suggestion, we further evaluated the token-level recovery accuracy of the restoration head on unperturbed prompts from CoPro and UnsafeBench. Exact Match measures whether the predicted token matches the original token, while LLM-based Semantic Match evaluates whether the predicted token is semantically consistent with the original token in the context of the sentence with GPT-5.2.

Token restoration performance on unperturbed prompts from CoPro and UnsafeBench (%).

EvaluationTop-1Top-5
Exact Match29.045.0
LLM-based Semantic Match44.082.0

Semantic matching yields higher recovery rates than exact token matching, suggesting that the restoration head can predict contextually relevant tokens even when the original tokens are not recovered exactly. The top-five results further indicate that semantically relevant alternatives are frequently among the leading candidate tokens. These results suggest that the restoration head provides useful semantic information for subsequent explicit visual grounding. We have added these results and further discussion in Appendix C (L964-971).

AuthorsOpenReview ↗

Response to comments on additional target model, safety assessment, and contextual rank analysis

W5: In Section 4.1, the evaluation uses one target model, its safety checker as ground truth, and one LLM judge whose identity is not stated, with no confidence intervals on gaps of a few prompts. The authors should add a second target model and classifier.

Q2: Which LLM serves as the final judge?

To address this concern, we further evaluated EEGIS on a second target model, SDXL 1.0. We include ESD, SLD, SAFREE, and Latent Guard, which were included in our original experiments, as baselines. MACE and RECE were excluded as their official SDXL implementations and checkpoints were unavailable, while adapting GuardT2I to SDXL would require retraining its conditional LLM (training on approximately 10 million prompt–embedding pairs, requiring 768 GPU-hours [1]).

Attack Success Rates (ASR, ↓)

MethodSneakyPromptMMA-DiffusionRing-A-BellP4DPGJAVG. (↓)
ESD57.9470.5999.4582.3750.9172.25
SLD-Medium37.1460.2297.7971.5929.8359.31
SLD-Strong29.7152.4991.7162.9024.8652.33
SLD-Max30.8648.6291.7162.9024.8651.79
SAFREE26.2940.8890.0653.3324.3146.97
Latent Guard18.8631.4944.7558.265.5231.78
EEGIS1.716.080.001.453.872.62

 

False Positive Rates (FPR, ↓)

MethodLAION-COCODiffusionDB
ESD2.753.75
SLD-Medium2.255.75
SLD-Strong1.257.00
SLD-Max2.756.50
SAFREE0.001.00
Latent Guard14.0018.50
EEGIS0.507.00

EEGIS achieves the lowest ASR across adversarial attack methods on SDXL 1.0. Notably, its defense performance remains comparable to that on SDv1.4, while maintaining similar FPRs. These results suggest that EEGIS maintains its defense effectiveness across different diffusion backbones. We have added these experimental results and discussion in Appendix C (L1057-1073).

In addition, we follow the established evaluation protocol using the stable diffusion safety checker, as described in Section 4.1. We also thank the reviewer for pointing out that the identity of the LLM judge was not stated. We used GPT-5.2 for the final LLM-based safety assessment and added this clarification in Section 4.1 (L357-358) and Appendix B (L845).

[1] Yang et al., GuardT2I: Defending Text-to-Image Models from Adversarial Prompts, NeurIPS 2024.

Q1: For each attack, what fraction of adversarial prompts contains at least one token above the threshold?

The proportions of adversarial prompts containing at least one token above the threshold are 62.6% for SneakyPrompt, 100% for MMA-Diffusion, 100% for Ring-A-Bell, 81.3% for P4D, and 20.0% for PGJ. These differences reflect how adversarial attacks vary in their degree of lexical perturbation. In particular, PGJ exhibits a much lower proportion, as it relies on lexically plausible substitutions. Importantly, EEGIS proceeds with explicit visual grounding even when no lexical anomalies are detected, allowing it to reveal harmful implications that remain concealed in lexically plausible expressions. This helps explain why EEGIS maintains low ASR across attacks despite differences in lexical anomaly rates.

Reviewer_aZQKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies defenses against adversarial attacks on text-to-image (T2I) models. The proposed EEGIS reveals the hidden intent of input prompts through three stages: lexical anomaly detection, semantic verbalization, and explicit visual grounding. The recovered and explicit prompts then support reliable safety assessment. The method is evaluated on NSFW-200 and T2I-RiskyPrompt across diverse attack methods and achieves stronger performance than existing baselines.

Critical strengths
  • EEGIS proposes and combines three stages to detect concealed harmful intent in adversarial prompts.
  • EEGIS achieves strong performance on the benchmarks compared to the baselines.
Critical weaknesses
  • The method is evaluated only on a single diffusion backbone (Stable Diffusion v1.4). It is unclear whether it generalizes to recent T2I architectures.

  • Architecture dependence of semantic verbalization. Semantic verbalization relies on intermediate U-Net features and token-wise cross-attention maps from early denoising steps. It therefore appears applicable only to U-Net-based diffusion models. Recent models such as SD3 and FLUX use DiT/MMDiT backbones with joint text–image attention and flow-matching objectives, and it is unclear how token-specific representations would be extracted from them.

  • The FPR of EEGIS is higher than that of several transformative defenses. This suggests a trade-off between lower ASR and higher FPR. The authors should analyze this trade-off, for example by reporting ASR and FPR under different detection thresholds.

  • The method depends heavily on GPT-5.2, which is used for explicit visual grounding. The authors should report results with open-source LLMs (e.g., Llama, Qwen) in place of GPT-5.2, to show that the method remains effective without a proprietary model. In addition, the cost analysis does not account for GPT-5.2. Calling a large proprietary LLM for every prompt likely incurs substantially higher computational costs than the baselines, most of which do not use such models.

Questions for authors
  • Is the proposed method applicable to recent backbones or other T2I models, such as DiT-based or flow-matching models (e.g., SD3, FLUX)?
3 comments
AuthorsOpenReview ↗

Response to comments on generalization to different diffusion backbones

We sincerely thank the reviewer for thoughtful feedback and the time devoted to reviewing our work. We have carefully considered the comments and hope that the additional experiments and explanations address the concerns raised.

In summary, we evaluated EEGIS on an additional T2I backbone (SDXL 1.0) and demonstrated robust defense performance with smaller open-source LLMs. We also discuss the applicability of our approach to other T2I architectures.

We provide detailed responses to each comment below.

W1: The method is evaluated only on a single diffusion backbone (Stable Diffusion v1.4). It is unclear whether it generalizes to recent T2I architectures.

To address this concern, we further evaluated EEGIS on SDXL 1.0, as shown below. We include ESD, SLD, SAFREE, and Latent Guard, which were included in our original experiments, as baselines. MACE and RECE were excluded as their official SDXL implementations and checkpoints were unavailable, while adapting GuardT2I to SDXL would require retraining its conditional LLM (training on approximately 10 million prompt–embedding pairs, requiring 768 GPU-hours [1]).

Attack Success Rates (ASR, ↓)

MethodSneakyPromptMMA-DiffusionRing-A-BellP4DPGJAVG. (↓)
ESD57.9470.5999.4582.3750.9172.25
SLD-Medium37.1460.2297.7971.5929.8359.31
SLD-Strong29.7152.4991.7162.9024.8652.33
SLD-Max30.8648.6291.7162.9026.8651.79
SAFREE26.2940.8890.0653.3324.3146.97
Latent Guard18.8631.4944.7558.265.5231.78
EEGIS1.716.080.001.453.872.62

 

False Positive Rates (FPR, ↓)

MethodLAION-COCODiffusionDB
ESD2.753.75
SLD-Medium2.255.75
SLD-Strong1.257.00
SLD-Max2.756.50
SAFREE0.001.00
Latent Guard14.0018.50
EEGIS0.507.00

As shown in the tables, EEGIS achieves the lowest ASR across adversarial attack methods on SDXL 1.0. Notably, its defense performance remains comparable to that on SDv1.4, while maintaining similar FPRs. These results suggest that EEGIS maintains its defense effectiveness across different diffusion backbones. We have added these experimental results and further discussion in Appendix C (L1057-1073).

We would like to note that evaluating defenses on a single diffusion backbone is common in prior T2I defense studies, including our baselines [1-5]. Our original evaluation followed this established setting, while the additional experiments further assess the generalizability of EEGIS across different diffusion backbones.

[1] Yang et al., GuardT2I: Defending Text-to-Image Models from Adversarial Prompts, NeurIPS 2024.

[2] Gandikota et al., Erasing Concepts from Diffusion Models, ICCV 2023.

[3] Gong et al., Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models, ECCV 2024.

[4] Lu et al., MACE: Mass Concept Erasure in Diffusion Models, CVPR 2024.

[5] Liu et al., Latent Guard: A Safety Framework for Text-to-Image Generation, ECCV 2024.

AuthorsOpenReview ↗

Response to comments on applicability of EEGIS to other T2I models and the ASR–FPR trade-off

W2: Architecture dependence of semantic verbalization. Semantic verbalization relies on intermediate U-Net features and token-wise cross-attention maps from early denoising steps. It therefore appears applicable only to U-Net-based diffusion models. Recent models such as SD3 and FLUX use DiT/MMDiT backbones with joint text–image attention and flow-matching objectives, and it is unclear how token-specific representations would be extracted from them.

Q1: Is the proposed method applicable to recent backbones or other T2I models, such as DiT-based or flow-matching models (e.g., SD3, FLUX)?

EEGIS was deliberately designed for U-Net-based Stable Diffusion models, which are widely adopted in the T2I defense literature and serve as the common evaluation setting for most of our baselines [1–5]. This allows for consistent comparisons with existing methods.

While our current implementation utilizes U-Net intermediate features and cross-attention maps, the underlying principle of semantic verbalization is not inherently limited to U-Net architectures. DiT/MMDiT-based models such as SD3 and FLUX also provide intermediate visual representations and joint text–image attention. In particular, token-specific representations could be constructed by extracting text–image attention weights from joint attention layers and using them to aggregate intermediate image-token features.

Although these models employ different architectures and flow-matching objectives, the availability of intermediate visual features and text–image interactions provides a potential basis for semantic verbalization. However, extending EEGIS to these architectures would require adapting the feature extraction and restoration modules to their architectural differences. We consider this an important direction for future work beyond the scope of the present study.

W3: The FPR of EEGIS is higher than that of several transformative defenses. This suggests a trade-off between lower ASR and higher FPR. The authors should analyze this trade-off, for example by reporting ASR and FPR under different detection thresholds.

Regarding the ASR–FPR trade-off, we would like to highlight the comparison between EEGIS and existing transformative defenses. Although some transformative defenses achieve lower FPRs, they exhibit an average ASR of 46.23%, approximately 20.6 times higher than EEGIS (2.24%), while their average FPRs remain comparable (2.04% vs. 1.50% on LAION-COCO; 5.71% vs. 6.50% on DiffusionDB), as shown in Tables 1 and 2. Also, we would like to clarify that EEGIS employs an LLM-as-a-Judge for binary safety assessment, supported by human verification (Appendix D), rather than applying a threshold to a continuous detection score. Nevertheless, following your suggestion, we further conducted a sensitivity analysis of the contextual rank threshold (τ\tau) to examine its impact on the ASR–FPR trade-off.

Attack Success Rates (ASR, ↓)

SneakyPromptMMA-DiffusionRing-A-BellP4DPGJ
τ=100\tau = 1002.865.520.551.744.42
τ=1000\tau = 10002.294.420.001.163.31
τ=5000\tau = 50002.295.520.001.164.42

 

False Positive Rates (FPR, ↓)

LAION-COCODiffusionDB
τ=100\tau = 1002.008.75
τ=1000\tau = 10001.506.50
τ=5000\tau = 50000.756.50

As shown in the tables, EEGIS maintains low ASR across all five attack methods at the evaluated thresholds, while FPR generally decreases as τ\tau increases. These results are consistent with the complementary roles of semantic verbalization and explicit visual grounding, which can reveal harmful implications even when no suspicious tokens are detected. Overall, EEGIS maintains effective defense performance across different threshold settings, providing a favorable balance between ASR and FPR. We have added these results and further discussion in Section 4.5 (L498-525).

AuthorsOpenReview ↗

Response to comments on experiments with open-source LLMs and API cost

W4: The method depends heavily on GPT-5.2, which is used for explicit visual grounding. The authors should report results with open-source LLMs (e.g., Llama, Qwen) in place of GPT-5.2, to show that the method remains effective without a proprietary model. In addition, the cost analysis does not account for GPT-5.2. Calling a large proprietary LLM for every prompt likely incurs substantially higher computational costs than the baselines, most of which do not use such models.

Following your suggestion, we further evaluated EEGIS by replacing GPT-5.2 with two open-source LLMs, Qwen-3.5-9B and GPT-OSS-20B, as shown in the tables below.

Attack Success Rates (ASR, ↓)

MethodSneakyPromptMMA-DiffusionRing-A-BellP4DPGJAVG. (↓)
ESD14.2927.6254.1447.5411.6031.04
SLD-Medium48.0074.5999.4582.9042.5469.50
SLD-Strong38.2964.6499.4580.5836.4663.88
SLD-Max26.8661.3391.7175.9424.3156.03
RECE17.7155.8039.7865.5121.5540.07
SAFREE14.8648.0797.7968.7018.7849.64
MACE12.5710.5013.8115.9414.3613.44
Latent Guard18.2930.9445.3058.844.4231.56
GuardT2I9.714.420.552.616.084.67
EEGIS (Qwen-3.5-9B)4.578.840.001.747.184.47
EEGIS (GPT-OSS-20B)4.009.940.001.455.524.18
EEGIS (GPT-5.2)2.294.420.001.163.312.24

 

False Positive Rates (FPR, ↓)

MethodLAION-COCODiffusionDB
ESD0.754.75
SLD-Medium2.255.75
SLD-Strong1.257.00
SLD-Max2.756.50
RECE3.003.75
SAFREE1.003.75
MACE3.258.50
Latent Guard14.0018.25
GuardT2I7.5053.00
EEGIS (Qwen-3.5-9B)0.003.25
EEGIS (GPT-OSS-20)0.005.00
EEGIS (GPT-5.2)1.506.50

Both open-source variants achieve relatively lower average ASR than GuardT2I, the strongest baseline, despite some performance degradation relative to GPT-5.2. Furthermore, they maintain low FPRs on both LAION-COCO and DiffusionDB, indicating that competitive defense performance can be achieved without substantially increasing false positives. These results suggest that EEGIS remains effective with smaller open-source LLMs, supporting its applicability beyond a specific proprietary model. We have added experimental results and further discussion in Appendix C (L1005-1013).

In addition, to address the computational cost concern, we measured the GPT-5.2 API cost incurred by EEGIS across adversarial prompts, as shown below.

API Cost of EEGIS with GPT-5.2 (USD per Prompt, ↓)

SneakyPromptMMA-DiffusionRing-A-BellP4DPGJAVG.
0.003380.003080.000340.002240.003450.00231

EEGIS incurs an average API cost of $0.00231 per prompt. The cost varies across attack methods, partly because prompts identified by early NSFW keyword filtering do not require subsequent LLM calls. This analysis quantifies the additional API cost associated with GPT-5.2, complementing the runtime analysis reported in our original evaluation. We have added these results and discussions in Appendix C (L959-962).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.