- 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
-
The proposed framework significantly enhances the interpretability of adversarial prompt defenses in T2I models by converting perturbed inputs into explicit visual descriptions.
-
By leveraging visually guided semantic restoration rather than static concept lists, the algorithm demonstrates strong adaptability to unseen malicious concepts.
Critical weaknesses
- The paper lacks ablation experiments on the key hyperparameter , and the authors do not analyze how ASR and FPR vary with different values of ; a clear quantitative guideline for setting across different T2I models should be provided.
- 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.
- The authors do not evaluate the framework against adaptive adversaries, leaving the worst-case performance of both proposed modules unverified.
Questions for authors
- 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)?
- How do ASR and FPR change when varying the hyperparameter ? Could you provide experimental results for different values of and explain how to choose this parameter for other T2I models?
3 comments
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, ↓)
| Method | SneakyPrompt | MMA-Diffusion | Ring-A-Bell | P4D | PGJ | AVG. (↓) |
|---|---|---|---|---|---|---|
| ESD | 14.29 | 27.62 | 54.14 | 47.54 | 11.60 | 31.04 |
| SLD-Medium | 48.00 | 74.59 | 99.45 | 82.90 | 42.54 | 69.50 |
| SLD-Strong | 38.29 | 64.64 | 99.45 | 80.58 | 36.46 | 63.88 |
| SLD-Max | 26.86 | 61.33 | 91.71 | 75.94 | 24.31 | 56.03 |
| RECE | 17.71 | 55.80 | 39.78 | 65.51 | 21.55 | 40.07 |
| SAFREE | 14.86 | 48.07 | 97.79 | 68.70 | 18.78 | 49.64 |
| MACE | 12.57 | 10.50 | 13.81 | 15.94 | 14.36 | 13.44 |
| Latent Guard | 18.29 | 30.94 | 45.30 | 58.84 | 4.42 | 31.56 |
| GuardT2I | 9.71 | 4.42 | 0.55 | 2.61 | 6.08 | 4.67 |
| EEGIS (Qwen-3.5-9B) | 4.57 | 8.84 | 0.00 | 1.74 | 7.18 | 4.47 |
| EEGIS (GPT-OSS-20B) | 4.00 | 9.94 | 0.00 | 1.45 | 5.52 | 4.18 |
| EEGIS (GPT-5.2) | 2.29 | 4.42 | 0.00 | 1.16 | 3.31 | 2.24 |
False Positive Rates (FPR, ↓)
| Method | LAION-COCO | DiffusionDB |
|---|---|---|
| ESD | 0.75 | 4.75 |
| SLD-Medium | 2.25 | 5.75 |
| SLD-Strong | 1.25 | 7.00 |
| SLD-Max | 2.75 | 6.50 |
| RECE | 3.00 | 3.75 |
| SAFREE | 1.00 | 3.75 |
| MACE | 3.25 | 8.50 |
| Latent Guard | 14.00 | 18.25 |
| GuardT2I | 7.50 | 53.00 |
| EEGIS (Qwen-3.5-9B) | 0.00 | 3.25 |
| EEGIS (GPT-OSS-20) | 0.00 | 5.00 |
| EEGIS (GPT-5.2) | 1.50 | 6.50 |
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, ↓)
| SneakyPrompt | MMA-Diffusion | Ring-A-Bell | P4D | PGJ | AVG. |
|---|---|---|---|---|---|
| 0.00338 | 0.00308 | 0.00034 | 0.00224 | 0.00345 | 0.00231 |
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).
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.
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.
We would like to clarify that EEGIS does not utilize a hyperparameter denoted by k. We understand the k mentioned in your comment to refer to the contextual rank threshold τ introduced in Section 3.1, which determines the tokens selected for semantic verbalization.
Following your suggestion, we conducted a sensitivity analysis of τ to examine its effects on ASR and FPR.
Attack Success Rates (ASR, ↓)
False Positive Rates (FPR, ↓)
As shown in the tables, EEGIS maintains low ASR across all five attack methods at the evaluated thresholds, while FPR generally decreases as τ 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.