acceptodds
Withdrawn from ICLR 2027

Immune Escape in Model Immunization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Efficient fine-tuning techniques have substantially reduced the cost of adapting open-weight models, enabling malicious actors to readily fine-tune them toward harmful capabilities. In response, model immunization modifies a pretrained checkpoint to resist subsequent adaptation toward designated harmful capabilities while preserving its utility and adaptability for benign tasks. However, we observe that existing immunization methods substantially overfit to the specific configurations used during immunization, particularly the images, textual prompts, and adaptation strategy the surrogate attacker is assumed to use. This fragility is obscured by existing evaluations, which largely reuse the same datasets and attack configurations employed during immunization. To rigorously evaluate model immunization, we develop ESCAPE, an attack framework that generates diverse variants of harmful adaptation designed to circumvent immunization. Specifically, ESCAPE misaligns the attacker from the configuration simulated during immunization along three axes: harmful data, model initialization, and adaptation strategy. We evaluate ESCAPE against three existing model immunization methods across diverse tasks and modalities, spanning text-to-image generation, image classification, and language models. Our results show that ESCAPE substantially weakens immunization across all three modalities and, in many cases, largely erases the protection that immunization provides over the un-immunized model.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
model immunization, harmful fine-tuning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_5tZ7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the vulnerability of model immunization, a defense designed to prevent open-weight models from being fine-tuned toward harmful tasks. The authors show that current immunization methods overfit to the defender's simulated attacker setup, creating an overlooked evaluation gap. To expose this fragility, they propose ESCAPE, a framework that intentionally departs from the defender's assumptions across harmful data, model initialization, and adaptation strategies like LoRA configurations. Empirical evaluations across three modalities (diffusion, classification, and language models) demonstrate that ESCAPE effectively bypasses existing defenses and substantially degrades their protection.

Critical strengths
  1. The paper identifies key limitations in current model immunity/safety evaluation protocols and proposes a comprehensive, multi-dimensional evasion framework.
  2. The authors conduct extensive experiments across diverse tasks, demonstrating the effectiveness and generalizability of the proposed evasion method.
Critical weaknesses

Lack of adaptive defense evaluation: The paper does not evaluate against an adaptive defender. Specifically, the authors do not report the attack success rate in worst-case scenarios where the defender has full knowledge of the attack algorithm.

Unclear novelty regarding the immunity vs. harmfulness adaptation gap: The paper frames the mismatch between model immunity defense and downstream harmfulness adaptation as a train-test discrepancy. However, it is unclear whether this work is the first to formalize this specific gap. The authors should provide a more thorough discussion of how prior safety benchmarks and fine-tuning protocols approached this problem to clarify the conceptual novelty.

Questions for authors
  1. Is this work the first to identify this discrepancy in model immunity evaluation, or do existing benchmarks also consider it? Please explicitly contrast your evaluation setup with prior protocols to clarify the core conceptual difference.
  2. How does the proposed bypass method perform if the defender has full knowledge of your algorithm (white-box/adaptive defense setting)? Could you provide concrete results or a worst-case analysis under this threat model?
Reviewer_eXzGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper focuses on the robustness and safety of existing model immunization techniques. It proposes a systematic framework, ESCAPE, to evaluate whether immunized models can maintain their protection against different malicious fine-tuning strategies.

Critical strengths
  1. The paper focuses on an intriguing and important research question.
  2. The paper investigates the model immunization problem in both diffusion models and LLMs.
Critical weaknesses

1. Limited Technical Novelty. The proposed ESCAPE framework mainly combines existing techniques, including dataset modifications, random noise, PGA, GCG, LoRA rank adjustments, and so on. After reading the paper, it feels like the main contribution is a systematic stress-testing framework for evaluating the robustness of existing model immunization methods, rather than a novel attack method.

2. Questionable Threat Model and Unrestricted Adaptation Settings. The authors show that increasing the LoRA rank from 4 to 32 weakens the effectiveness of model immunization. However, a higher LoRA rank naturally provides greater capacity for model fine-tuning. In fact, I believe the threat model should impose reasonable constraints on the attacker's training configurations/hyperparameters. Otherwise, one could theoretically pre-train the immunized model on an entirely new dataset, update all its parameters, and then perform malicious fine-tuning. Such an approach might also bypass immunization. Without reasonable restrictions on the attacker's adaptation capabilities, it seems extremely difficult for any model immunization method to provide meaningful protection.

3. Unclear Effectiveness of GCG in Specifically Bypassing Immunization. In Table 5, GCG increases the harmful response rate of the immunized models, but it also improves the harmful adaptation performance of the original, unimmunized models. Therefore, it is unclear whether GCG specifically bypasses the immunization mechanism or simply produces training data that makes harmful fine-tuning more effective for any model. I believe the authors should better distinguish between these two effects.

4. Limited Evaluation on Recent LLM Backbones. I would like to see evaluations on more recent and advanced models, such as Qwen3.5, to better demonstrate the generalizability of the proposed approach.

Questions for authors

Please refer to the weakness section.

Reviewer_TB89OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies whether model immunization remains effective when the attacker uses different configurations from those assumed during immunization. The authors propose ESCAPE, a framework that explores three directions: harmful-data misalignment, adaptation misalignment, and model initialization misalignment. They evaluate ESCAPE against three existing immunization methods across diffusion models, image classification, and language models. The results show that several simple changes to the attack configuration can substantially weaken existing defenses. The problem is interesting, and the experimental findings are useful. However, I have several concerns about the novelty of the framework, the experimental coverage, and the reliability of some evaluation settings. I believe these issues need to be better addressed to fully support the paper's conclusions.

Critical strengths
  1. Interesting research question. The paper examines whether model immunization can remain effective when the actual attacker differs from the one simulated during defense training. This is an important problem for open-weight models, where the defender cannot control how released models are subsequently fine-tuned.

  2. Useful experimental analysis. The authors evaluate three existing immunization methods across diffusion models, image classification, and language models. The experiments also include different attack settings, matched random-perturbation controls, and absolute performance measurements. These results provide useful evidence of the limitations of existing immunization methods.

Critical weaknesses
  1. Relation to prior work. The paper argues that existing evaluations largely reuse the attack configurations used during immunization. However, IMMA already tests crossed adaptation between DreamBooth and Textual Inversion and uses separately generated images for immunization and attack. CLAMP also evaluates longer fine-tuning than the simulated horizon. These works have already considered different forms of attacker misalignment. What does ESCAPE add beyond testing more attack configurations?

  2. Incomplete evaluation. ESCAPE introduces three misalignment dimensions, but their experimental coverage is uneven. For example, initialization misalignment is only evaluated on CLAMP with Van Gogh, while the Combined attack only considers data-related modifications. What happens when all three dimensions are applied together? I would also like to see whether adaptation and initialization misalignment remain effective across other models and tasks.

  3. Questionable prompt misalignment evaluation. In Appendix E, the authors replace the original SG-Bench instruction with a forced-choice prompt that only allows the model to output A or B. However, selecting the harmful option does not necessarily mean the model has recovered the ability to generate harmful content. Could the authors evaluate whether models fine-tuned using these forced-choice prompts can also generate harmful responses in a free-form setting?

  4. Single safety judge. The main free-form LLM experiments rely on Llama-Guard-3-1B as the only safety judge. Given the large improvements reported under GCG, I would like to see whether these results hold with other safety judges or human evaluation. How reliable are the reported harmful response rates when evaluated by a different judge?

  5. Limited defense coverage. The experiments only consider IMMA, CondNum, and CLAMP. In particular, the reported classification results focus on CondNum, while the language model experiments only evaluate CLAMP. How well does ESCAPE generalize to other immunization methods, especially those designed to handle different adaptation strategies?

Questions for authors

Please refer to the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.