acceptodds
Under review as a conference paper at ICLR 2027

GARF: Gradient-Aligned Reactivation Forgetting for Backdoor Defense

Abstract

Backdoor defenses often suppress visible trigger behavior, yet leave latent backdoor features that can be reactivated by small, adversarially optimized perturbations. We study this problem in a realistic post-training setting where the defender only has access to a possibly backdoored model, without any poisoned data. We propose Gradient-Aligned Reactivation Forgetting (GARF), a bilevel post-training defense that treats the parameter difference between the original backdoored model and its defended counterpart as a proxy backdoor task direction in parameter space. In the lower level, GARF learns a small input-agnostic perturbation on clean data whose induced parameter gradients are encouraged to align with this direction, thereby simulating an adaptive reactivation attack using only accessible information. In the upper level, the model is updated to jointly reduce the clean loss and a reactivation loss evaluated under the learned perturbation, forming an iterative attack–defense process that repeatedly exposes and suppresses reactivation-sensitive directions. Empirically, GARF reduces the white-box reactivation attack success rate by over 50% on average and keeps the black-box reactivation ASR below 10%, while preserving competitive clean accuracy across multiple benchmark datasets and diverse attack types.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_oPKqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Current backdoor defenses leave latent backdoor features in the sanitized models, which allows the reactivation of the backdoor. To achieve deep erasure of the backdoor, the authors propose GRAF (Gradient-Aligned Reactivation Forgetting) as a defense that introduces a small input-agnostic perturbation to simulate an adaptive reactivation attack. Learning to lower the error loss against this perturbation enables clean alignment and gradient-level erasing of the reactivation backdoor attack.

Critical strengths
  1. Defending the reactivation backdoor is crucial.

The targeted defense objective is crucial to truly removing the implanted backdoor. The latent reactivation attack requires deep sanitization.

  1. Comprehensive evaluation.

The authors provide extensive experiments across related defenses and the proposed work, offering a comprehensive evaluation and showing the advantage of GARF.

Critical weaknesses
  1. Mismatch between the threat model and the GA perturbation

According to the abstract and the threat model, the defender has no prior knowledge of any potential backdoor. However, the construction of GA perturbation δGA\delta^{GA} uses the backdoor's target yty_t. This either violates the threat model, or the determination process of the backdoor's target is missing.

  1. The rationale of using the gradients ΔθL(θ,;x+δ,yt)\Delta_{\theta} L(\theta,; x+\delta, y_t) for the similarity measurement.

Within A(⋅)A(\cdot) in Eq(7), the second term measures the inner-product similarity between gθg_{\theta} and θb−θ\theta_b - \theta. However, there is a lack of an interpretation of why gθg_{\theta} can represent the gradient discrepancy θb−θ\theta_b - \theta.

  1. The objective loss function used in Eq(7) is missing.

The applied loss function for the perturbation generation is not introduced after Eq(7).

  1. Evaluation with entirely clean models is missing

As the threat model includes the situation where the defender may receive a clean, not backdoored model, the impact of using GARF is not investigated.

Questions for authors

See weaknesses above.

Reviewer_4GYzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper tackles backdoor reactivation, where a defended model's backdoor can be revived by a small adversarial perturbation. The proposed method, GARF, fine-tunes a backdoored model using only a small clean dataset: it searches for perturbations that push the model back toward its original backdoored parameters and then trains against them. The target label is obtained from an ensemble of existing detectors. Experiments across several datasets, architectures and attacks report much lower reactivation success than existing defenses while maintaining clean accuracy.

Critical strengths
  1. Post-defense reactivation is under-addressed, and using the displacement θb−θ\theta_b - \theta as a proxy backdoor direction is a reusable idea rather than another fine-tuning recipe.

  2. The paper provides extensive experiments and ablations. The reductions in reactivation ASR are large and consistent in direction.

Critical weaknesses
  1. GARF's inner problem is conditioned on yty_t, supplied by an ensemble of four detectors. The ensemble is therefore critical to the method, yet it is validated on only three attacks on CIFAR-10, and its accuracy under reduced data and across multiple seeds is not reported. Given how critical this requirement is, ensemble performance should be reported for every tested setting. Every result also appears to be a single run on a single clean subset, with no seeds or variance reported, so the stability of neither the detection nor the defense across random subsets can be judged. Finally, SAU and I-BAU are run untargeted while GARF uses yty_t. For a fair comparison, SAU and I-BAU should be evaluated with the target class provided; their inner optimisation can easily be adjusted to do so.

  2. GARF's inner subproblem and the reactivation attack share the same construction: a universal, yty_t-targeted ℓ∞\ell_\infty perturbation at the same budget. The reported gains are therefore consistent with adversarial training against the exact evaluation perturbation class. Since a central claim is that the proxy direction supplies the robustness, the result should survive switching the inner adversary to per-sample perturbations or to untargeted PGD. Neither is tested.

Questions for authors
  1. Can you report robust accuracy (RA), as in BackdoorBench? Low ASR does not by itself imply poisoned inputs are recovered to their correct class, so RA is needed to interpret the results.

  2. If SAU and I-BAU are run with their inner adversary targeted to the detected yty_t, at the same budget and on the same clean subset, does GARF's margin persist?

  3. What is the ensemble's detection accuracy in every tested setting (all attacks, datasets and data regimes)?

  4. How stable are the results, and the ensemble's detections, across multiple seeds and random clean subsets, including the reduced-data settings?

  5. What happens when the inner targeted objective is replaced with untargeted PGD (cross-entropy maximisation on the true label)? How much does reactivation robustness degrade? This would test whether alignment works only because the perturbations are targeted.

  6. If the universal δ\delta is replaced with per-sample perturbations in the inner problem at the same budget, and evaluated against the same universal reactivation attack, does the result hold? It should, if the proxy direction is what supplies the robustness.

  7. Does the cost reported in Tab. 10 include running the entire ensemble?

Reviewer_c9tYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Gradient Alignment Reactivation Forgetting (GARF) to prevent backdoor reactivation. Specifically, GARF looks at the parameter difference between the initial backdoor model and the current optimized model as a proxy for the backdoor direction. The inner optimization focuses on searching for perturbations on clean inputs while encouraging the backdoor loss descent direction to align with this proxy. The outer optimization updates the model using clean and perturbed inputs with their ground truth labels. Experiments show a substantial reduction in reactivation attack success rates across different attacks.

Critical strengths
  1. The paper focuses on a limitation of current backdoor defenses: Low attack success under the original trigger does not necessarily indicate that the backdoors have been removed. Therefore, a good defense should also consider preventing such reactivation. This motivation sounds good.

  2. The proposed method achieves good results in preventing backdoor reactivations. The experiments are sufficient, involving different attack scenarios and ablation studies on the clean data size, λ\lambda, and β\beta.

  3. The paper is organized well, and the presentation is good.

Critical weaknesses

I would give this paper a weak rejection at the review phase; However, I am open to further discussion during the rebuttal. I could improve my rating if the author can address my concerns.

  1. Technically, the paper involves two stages in backdoor reactivation defense: 1) target label detection (using NC, Unicorn, and DISTIL); 2) target label protection(GARF). However, the paper does not clearly describe the key experimental setup for target label detection. Since the defender does not know the trigger information or the targeted classes, the key experimental setup for target label detection must ensure that it can detect target classes as much as possible. Although the paper discusses multi-trigger and multi-targeted-class in Appendix E.2, it seems the main experiment in Table 2 mainly focuses on a single targeted-class scenario, which means the paper only considers the most suspicious class as the targeted class when doing target label detection. So, one of my major concerns is whether GARF remains effective when the detection stage selects an incorrect label or misses one of several actual targets. Besides, the computational cost of the complete defense pipeline is still not clear. As the paper applies three different target label detection approaches, the detection time may significantly increase when the task gets complex (consider a 100-class classification task). Also, one minor concern: GARF is limited to the image classification task. I am not sure if it can be generalized to text generation or multi-modality tasks.

  2. My biggest concern is the gradient alignment term in GARF. Using the displacement θb−θ\theta_b - \theta as the proxy backdoor direction is not convincing to me. The interpolation experiment shows that moving the defended model toward the original backdoored model can restore backdoor behavior. However, this does not directly establish that aligning perturbation-induced parameter descent directions with (θb−θ\theta_b-\theta) helps identify inputs that specifically reactivate the backdoor. The displacement may also reflect changes in benign features and general adversarial robustness.

Questions for authors

Q1: Does the author think the GAFR is still effective when the detection stage selects an incorrect label or misses one of several actual targets?

Q2: If the task gets complex, for example, a 100- or 1000-class classification task, do you think the target label detection is still efficient? If not, do you think GARF is limited to easy classification tasks?

Q3: Can you give me a detailed explanation of why aligning the perturbation-induced parameter descent directions with (θb−θ\theta_b-\theta) helps identify inputs that specifically reactivate the backdoor? Although the experiments show the effectiveness of GARF, I still cannot figure out the reason.

Q4: I am not sure if GARF would have a side effect such as trigger shifting Don't shift the trigger: Robust Gradient Ascent for Backdoor Unlearning. It looks like δ\delta is decoupled from the original backdoor label yty_t, but whether it would reassociate with its true label and form a clean-label backdoor attack is unknown. That means δ\delta will correlate with its ground-truth label, for example, dog. And any image patched with such δ\delta would be redirected to dog. So, my question is: when the clean defense data are class-imbalanced or omit certain classes, could GARF associate its generated perturbations with overrepresented classes, introducing new targeted misclassification behavior?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.