acceptodds
Under review as a conference paper at ICLR 2027

Trigger Isolation: Decomposing Triggers and Targets for LLM Backdoor Purification

Abstract

Backdoor purification aims to remove attacker-specified behaviors activated by hidden triggers while preserving benign model utility. As the trigger is typically unknown, recent defenses introduce defender-controlled proxy backdoors that associate chosen triggers with targets. This training requires the model both to learn the target and to condition its generation on the trigger. Because existing proxy constructions change both trigger presence and target supervision relative to their controls, the resulting parameter changes mix these two effects. As a result, purification may be sensitive to the targets associated with both the defender's proxies and the model's original backdoor. To verify this conjecture, we conduct controlled experiments, showing that purification effectiveness is unstable across targets, even when triggers are held fixed. This previously overlooked dependence motivates us to decompose proxy-induced parameter changes into trigger-conditioned and target-learning contributions. We propose Trigger Isolation (T-ISO), which contrasts two proxy models trained on the same prompts and targets, differing only in whether the trigger is present. This matched comparison reduces shared target-learning effects and better isolates trigger-conditioned parameter changes to guide purification. We evaluate T-ISO across multiple models, attacks, and targets. T-ISO achieves the strongest overall performance among the evaluated defenses, particularly on strong backdoors. Across the same four Llama-2-7B-chat targets, it reduces mean post-recovery ASR from 18.50% to 3.50% under full-parameter defense and from 4.13% to 0.13% under LoRA-adapter defense. Its advantage persists across different victim strengths, showing that controlling target enables more effective removal of unknown LLM backdoors.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Backdoor Attacks, Backdoor Purification, Trigger–Target Disentanglement, Model Security

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_kpVUOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes T-ISO for purifying backdoors embedded in model weights. Its distinctive design constructs two proxy models and identifies shared sensitive channels through parameter differencing and cross-proxy ranking, followed by channel zeroing and fine-tuning on benign responses to remove unknown backdoors. Experimental results demonstrate that the proposed T-ISO algorithm achieves strong backdoor defense performance.

Critical strengths
  1. The manuscript presents a clear research motivation, and the proposed algorithm is novel.

  2. The design of the proposed T-ISO algorithm is reasonable and theoretically grounded.

  3. The experimental results demonstrate that T-ISO achieves strong defensive performance.

Critical weaknesses
  1. The poisoning ratio of 50% used in the manuscript is excessively high. Although an ablation study on the poisoning ratio is provided, such a substantial level of poisoning may potentially affect the validity of the analysis.

  2. The Triggered Branch of T-ISO is trained using 100% triggered samples, without any clean samples as a constraint. How can the authors ensure that the model has genuinely learned the backdoor trigger relationship, rather than simply tending to generate the target regardless of whether the trigger is present?

  3. T-ISO assumes that different triggers may rely on a shared internal mechanism. However, the possibility that the proxy backdoor and the original backdoor operate through non-overlapping internal pathways requires further verification.

  4. The authors should explicitly report the specific forms of the six proxy triggers.

  5. The manuscript provides relatively comprehensive validation of target generalization, but the evaluation of trigger generalization, particularly generalization across different trigger mechanisms, remains insufficient.

  6. The manuscript only compares proxy training time with Purify, without reporting the runtime, computational resource consumption, or related efficiency metrics for the other major baselines.

Questions for authors

Please refer to weaknesses.

Reviewer_E97SOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
2
Presentation
2(−1)
Soundness
3
Read review
Paper summary

This paper proposes Trigger Isolation (T-ISO), a method for purifying backdoored language models without knowing the attacker’s trigger or target or requiring a trusted clean model. Existing proxy-based defenses can mix changes associated with learning a target response and changes associated with trigger conditioning, making purification sensitive to target selection. T-ISO addresses this by training paired model copies on identical prompts and target responses, differing only in trigger presence. Their parameter differences reduce shared target-learning effects and guide the selection of feed-forward channels for suppression. Subsequent training on proxy-triggered prompts paired with benign answers helps restore normal behavior.

Critical strengths

1. The controlled experiments provide a convincing diagnosis of target sensitivity.
Figure 1(c) and Appendix B vary the victim target while keeping the trigger fixed, and separately vary proxy targets while keeping the victim fixed. These comparisons directly demonstrate that target choice affects purification effectiveness. The authors also explicitly acknowledge that these experiments do not establish the parameter-level cause, keeping their interpretation appropriately limited.

S2. The method is simple and directly addresses the identified problem.
Training the control branch on the same target response provides a straightforward way to reduce shared target-learning effects in the parameter contrast. The construction also processes fewer training records than Purify (2KEn2KEn versus 3KEn3KEn), while consensus ranking avoids pairwise cosine comparisons between proxies. These are concrete computational advantages, although they do not establish an end-to-end runtime improvement.

S3. The evaluation covers a broad range of settings.
The experiments include two models, five attacks, two tasks, full-parameter and LoRA defenses, different target responses, and variations in poisoning ratio and initial attack strength. Table 2 is particularly informative: under full-parameter defense, T-ISO’s ASR ranges from 0.50% to 6.50% across targets, compared with 10.50% to 24.50% for Purify. The narrower range supports improved consistency in addition to lower average ASR.

S4. The paper clearly reports unfavorable results and experimental limitations.
Appendix C acknowledges that T-ISO performs worse than Purify immediately after suppression on Qwen3’s LoRA sentiment task. The authors also disclose that runtime measurements are not hardware-matched and report substantial residual ASR in the component ablations. These disclosures help readers distinguish the effectiveness of individual components from that of the complete pipeline.

Critical weaknesses

1. The missing recovery-only control limits attribution of the overall improvement.
Table 3(b) shows that proxy-triggered recovery reduces ASR from 53.25% to 7.38%, whereas clean recovery from the same suppression checkpoint achieves 47.25%. This establishes the importance of the recovery procedure, but the paper does not evaluate proxy-triggered recovery without preceding suppression. Consequently, it remains unclear how much the proposed signature construction contributes beyond this recovery procedure alone. Clean fine-tuning is not an adequate substitute for this control because it uses different inputs.

2. The ablations demonstrate an interaction between components but provide limited evidence for the proposed mechanism.
In Table 3(a), replacing only the contrast reduces suppression ASR slightly, from 98.63% to 97.19%, while replacing only the scoring rule yields 98.88%. The substantial improvement appears only when the matched contrast and consensus ranking are combined, reaching 53.25%. These results support the effectiveness of the combination, but do not establish that removing shared target-learning effects is the mechanism responsible for the improvement.

3. The practical relevance of the local theoretical analysis remains unverified.
Theorem 1 provides a first-order characterization under full-batch gradient descent, with an error bound εS=βGT2(1−1/S)\varepsilon_S=\beta GT^2(1-1/S), where T=ηST=\eta S. However, the experiments train proxy branches for multiple epochs, and the paper does not estimate this error or check whether the channel-ranking condition remains informative under the actual training configuration. The bound is not necessarily invalid, but its usefulness for explaining the implemented method is unestablished.

The authors appropriately acknowledge that the analysis does not guarantee transfer to unknown backdoors or successful purification. Nevertheless, experiments connecting the local analysis to the learned signatures would strengthen the explanation. For example, the paper could compare residual directions, channel rankings, and purification outcomes across different proxy-training durations, including a single-step setting. Exact cancellation at one step does not guarantee effective purification, but such measurements would test whether the proposed first-order explanation remains relevant as training proceeds.

4. Robustness against attacks designed to evade the defense is not evaluated.
The paper explicitly leaves broader adaptive-attack evaluation to future work. This is a significant limitation because the defense makes structural choices that an informed attacker could potentially exploit. Channel suppression is restricted to MLP projections, leaving open whether backdoors trained primarily through attention parameters remain effective. Likewise, minimum-percentile aggregation can assign a low score to a channel that ranks poorly under even one proxy, motivating attacks that deliberately disrupt cross-proxy consistency.

5. Statistical reporting is insufficient to assess small differences and the stability of the motivating observations.
The paper does not report results across multiple random seeds, confidence intervals, or variability across repeated runs. In evaluations using 200 prompts, ASR differences of 0.50 percentage points correspond to a single example, so small numerical advantages should not be treated as established improvements without uncertainty estimates. “Zero observed ASR” is appropriately qualified, but reporting success counts and confidence bounds would make its meaning clearer. For example, zero successes among 200 independent trials gives an approximate one-sided 95% upper bound of 1.5%; refusal evaluations should instead use their actual scored denominators.

The same concern applies to the motivation experiments. Figure 1(c) and Appendix B provide useful controlled comparisons, but they use one model and one victim trigger without reported variation across training runs. Repeating these comparisons across seeds and triggers would establish whether the observed target sensitivity is systematic and how its magnitude compares with ordinary training variability.

Questions for authors

N/A, see the weakness.

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

This paper identifies a limitation of existing proxy-based backdoor purification methods, such as Purify, which locate backdoor-related channels by comparing the parameter differences between proxy-backdoored models and benignly fine-tuned models. However, these differences capture both trigger conditioning and target learning, making purification effectiveness sensitive to the choice of attack targets. To address this issue, the authors propose Trigger Isolation (T-ISO). The method constructs two proxy training datasets with identical target outputs but differing in trigger presence, thereby reducing the influence of target learning on parameter differences. It then identifies backdoor-related channels through cross-proxy consensus ranking, followed by channel suppression and benign recovery fine-tuning. The authors further provide a first-order gradient analysis to justify the design and evaluate its effectiveness across different models, backdoor attacks, and attack targets.

Critical strengths

S1. Clear motivation that reveals an important limitation of existing proxy-based backdoor purification methods.

Through controlled experiments, the authors show that changing either the attack target or the proxy target can significantly affect purification effectiveness, even when the trigger remains fixed. This observation highlights the target dependence of existing proxy-based methods and provides a valuable perspective for improving this line of defense.

S2. A reasonable method with theoretical justification.

By matching target supervision, T-ISO shifts the focus of parameter comparison from joint changes in triggers and targets to trigger-conditioned changes. This design is reasonable compared with directly contrasting proxy-backdoored and benignly fine-tuned models. Furthermore, the authors provide a first-order gradient analysis to explain the cancellation of shared target-learning terms and extend the discussion to LoRA parameterization, offering preliminary theoretical support for the proposed method.

S3. Promising experimental results on backdoor purification.

The authors evaluate multiple backdoor attacks on Llama-2-7B-chat and Qwen3-8B under both full-parameter and LoRA defense settings. The results show that T-ISO achieves low attack success rates (ASRs) in most settings, with particularly consistent performance across different attack targets. The paper also includes ablation studies and robustness evaluations under varying poisoning ratios and initial attack strengths, providing reasonably comprehensive experimental evidence.

Critical weaknesses

W1. The necessity and unique advantages of the proposed method over existing backdoor purification approaches remain unclear.

The paper primarily compares T-ISO with Purify and conventional defenses such as fine-tuning and pruning. However, existing studies, such as CROW and LETHE, can also purify backdoored LLMs without knowing the backdoor triggers or requiring clean reference models. Therefore, T-ISO does not appear to offer a substantially less restrictive threat model. Meanwhile, T-ISO requires additional training of multiple proxy models, potentially introducing considerable computational overhead compared with approaches that do not rely on proxy backdoors. I believe this somewhat weakens the practical significance of the proposed method.

W2. Insufficient evidence of advantages in benign performance preservation, and an unclear trade-off between computational cost and defense effectiveness.

T-ISO purifies backdoored models by suppressing a relatively large proportion of model channels, which may affect their benign capabilities. However, the benign performance evaluation covers only one victim model setting, making it difficult to assess the method's ability to preserve benign performance under different backdoor conditions. Furthermore, Table 7 shows that T-ISO achieves essentially the same average benign performance as Purify, without a clear advantage, while some tasks exhibit noticeable performance degradation. For example, under the LoRA setting, GSM8K accuracy decreases from 11.14% for the undefended model to 5.84% after applying T-ISO. Therefore, although T-ISO achieves lower ASR, it remains unclear whether it can achieve the same ASR with less degradation in benign performance. Given the additional training overhead, a more systematic analysis of the trade-off among ASR, benign performance, and computational cost is necessary.

W3. The theoretical analysis and experimental results do not sufficiently explain the sources of performance improvements.

The theoretical analysis mainly demonstrates that matched-target supervision can eliminate shared target-learning terms under a first-order approximation. However, it does not sufficiently explain why parameter changes induced by proxy triggers can effectively identify the channels associated with unknown backdoors. Furthermore, compared with Purify, T-ISO simultaneously modifies the parameter contrast construction, channel aggregation strategy, and recovery fine-tuning procedure, making it difficult to attribute the observed improvements to a specific component. For example, Table 3 shows that using either the T-ISO contrast or consensus ranking alone still results in poor defense performance, while combining them substantially reduces ASR. This suggests that the method relies heavily on interactions among multiple components, but the paper lacks an in-depth analysis of this phenomenon. Consequently, it remains unclear how much of the performance improvement can be attributed specifically to the proposed Trigger Isolation mechanism.

Questions for authors

Q1. Necessity of the proposed approach and comparison with alternative methods.

Existing studies, such as CROW and LETHE, can already achieve backdoor purification without knowing the triggers or requiring clean reference models. What are the primary advantages of T-ISO over these approaches that do not require proxy backdoor training? Could the authors compare ASR, benign performance, and GPU time under the same models, attacks, and benign data budgets to justify the necessity of additional proxy training?

Q2. Benign performance preservation.

Table 7 shows that T-ISO achieves average benign performance comparable to Purify, while some tasks exhibit noticeable degradation. Could the authors evaluate benign performance across more victim models and backdoor attack settings? In addition, I suggest varying hyperparameters such as the channel suppression ratio and comparing the benign performance of different methods at matched ASR levels. This would provide a more accurate assessment of the trade-off between defense effectiveness and model utility.

2 comments

Thank you to the authors for the detailed rebuttal and extensive additional experiments. I appreciate that the new controlled ablations provide stronger evidence for the independent contribution of target matching and channel selection, while the expanded utility and runtime evaluations substantially address my concerns regarding the effectiveness–cost trade-off. These results strengthen the empirical justification for T-ISO. However, some concerns remain, particularly the limited mechanistic understanding of why proxy-derived signatures generalize to unknown backdoors, and the absence of comprehensive adaptive-attack evaluations.

Overall, I acknowledge the substantial improvements made during the rebuttal. I therefore decide to maintain my score of 3 (since it is already positive) and encourage the AC to consider both the strengthened empirical evidence and the remaining limitations in the final decision.

We sincerely thank the reviewer for the continued positive assessment and recognition of our additional experiments. We are glad that the controlled ablations and expanded utility and runtime evaluations strengthen the empirical support for T-ISO. We are also conducting adaptive-attack experiments under two settings and will share the results once they are complete. Thank you again for your thoughtful evaluation and constructive feedback.

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

This paper studies how to remove unknown backdoors from LLMs using defender-controlled proxy backdoors. Its main observation is that comparing a proxy-backdoor model with a benign-trained control changes both trigger presence and target supervision, so the resulting parameter difference can reflect both effects.

The proposed method, T-ISO, trains two proxy branches on the same prompts and target responses, differing only in whether the trigger is present. It aggregates channel-level differences across proxies using consensus ranking, suppresses selected channels in the original suspicious model, and then performs recovery training on proxy-triggered prompts with benign answers. A local analysis explains the first-order effect of matching target supervision. Experiments cover two model families, several attacks, different target responses, and both full-parameter and adapter-only defenses.

Critical strengths
  1. The matched-target control is a sensible improvement to proxy-based purification. Holding the response target fixed makes the comparison easier to interpret than changing both the trigger and the supervision. The paper also avoids claiming that trigger and target mechanisms occupy separate parameter sets. Its theory explicitly describes a local, first-order cancellation and leaves transfer to unknown backdoors to empirical evaluation. That is an appropriate distinction.

  2. The reported improvements are substantial across several settings, and the ablations provide useful evidence. For example, Table 2 reduces mean post-recovery ASR from 18.50% to 3.50% under full-parameter defense. Table 3 also tests contrast construction and aggregation separately: neither change alone works well on those four victims, while their combination substantially reduces suppression-only ASR. I appreciate that the appendix reports cases where suppression alone is less successful, rather than presenting the method as uniformly effective.

Critical weaknesses
  1. The final gains are not yet cleanly attributed to the proposed signature rather than the recovery procedure. T-ISO changes the proxy data, contrast, aggregation rule, and recovery inputs. Table 3 makes progress on separating these changes, but it also shows that recovery is crucial: after identical T-ISO suppression, clean recovery leaves 47.25% mean ASR, while proxy-triggered recovery reduces it to 7.38%.

    The missing comparisons are proxy-triggered recovery without suppression, and the same recovery applied after Purify or random suppression. The existing FT-only and random-suppression controls use clean recovery, so they do not answer this question. In addition, the proxy constructions use different data compositions and training budgets. A budget-matched comparison would make the evidence for target matching much stronger.

  2. The evaluation needs stronger checks before low ASR can be interpreted as successful purification. The paper uses a deterministic prefix-match criterion, and the refusal evaluation scores nonempty responses. Appendix D.1 already notes that low ASR can accompany empty outputs. This raises a practical concern: a defense may alter the output format, move the target later in the response, or stop generating useful answers without removing the unwanted behavior. Utility is evaluated on only one Llama-2 BadNets setting, so it does not resolve this concern across the stronger Qwen3 experiments and ablations. Even there, the LoRA GSM8K score drops from 11.14 to 5.84 despite an improved overall average.

    There are also unexplained differences between raw ASRs in tables described as covering the same victims and protocols. For example, the LoRA code-target victim has 5.00% raw ASR in Table 2 but 21.50% in Table 5. The full-parameter hostility victim has 52.00% in Table 2 but 55.00% in Table 5. These may reflect different checkpoints or evaluation revisions, but the current text does not explain them. This matters when using the supplementary results to assess what suppression achieves before recovery.

Questions for authors
  1. How well does proxy-triggered benign recovery work on its own, without channel suppression? What happens when the same recovery is applied after Purify and random suppression, using the same records and training budget?

  2. Can you reconcile the raw ASR differences between Tables 2 and 5? Please identify the victim checkpoint, prompt set, decoding configuration, and scoring rule for each table, and clarify whether the suppression-only and recovered results form matched comparisons.

  3. Could you report empty-response rates and useful-answer rates alongside ASR, particularly for refusal attacks and the strong Qwen3 victims? A semantic assessment on a representative subset would also help establish whether failures of exact prefix matching correspond to removal of the attack behavior.

  4. How stable are the results across independently trained victims and different proxy panels? The current proxy-count study uses one victim and nested subsets of one panel. Repeated runs and utility measurements for the main ablation settings would help show that the signature is reliable rather than specific to the selected proxies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.