acceptodds
Withdrawn from ICLR 2027

Distribution Aligned Reweighting for Backdoor Defense

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Backdoor attacks pose a serious threat to deep neural networks by embedding a backdoor to induce malicious behavior on triggered inputs while preserving normal performance on clean inputs. Existing in-training defenses often rely on assumptions about differences in the learning dynamics of clean and poisoned samples, which may not hold consistently across attacks. We propose Distribution-Aligned Reweighting (DAR), an in-training backdoor defense that views backdoor poisoning as an adversarial shift of the training distribution away from the underlying clean distribution. Given a small trusted clean reference set, DAR performs class-wise distribution alignment to assign weights to training samples and uses the resulting weights to construct a sampling distribution for standard model training. DAR does not require prior knowledge of the attack and does not modify the model architecture or training objective. Extensive experiments across CIFAR-10, GTSRB, and ImageNet, covering diverse backdoor attacks and poisoning ratios, demonstrate that DAR consistently achieves low attack success rates while maintaining high clean accuracy and strong robust accuracy on triggered inputs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor Defense, Trustworthy Machine Learning, Robust Training, Distribution Shift

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 4 reviews, above 27% of the others.
Reviewer_j6DAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Distribution-Aligned Reweighting (DAR), an in-training backdoor defense method that views backdoor poisoning as an adversarial distribution shift between the potentially poisoned training data and the underlying clean data. Unlike existing defenses that rely on differences in the learning dynamics of clean and poisoned samples, DAR leverages a small trusted clean reference set to guide sample reweighting. Specifically, the method first extracts feature representations of the training and reference samples, and then applies class-wise Kernel Mean Matching (KMM) to assign sample weights by minimizing the distribution discrepancy between the weighted training data and the clean reference data. The key intuition is to suppress or exclude samples that contribute to the discrepancy from the clean distribution. Under the overall KMM distribution-matching objective, samples assigned low weights are considered less useful for aligning the training distribution with the clean reference distribution.

Critical strengths
  • The paper provides an intuitive perspective on backdoor defense by treating data poisoning as a distribution shift and mitigating it through distribution alignment.

  • DAR is straightforward to implement and does not require modifications to the model architecture or training objective.

  • The method achieves low attack success rates and competitive robust accuracy across multiple datasets, attacks, and poisoning ratios.

Critical weaknesses
  • Sample reweighting techniques for backdoor defense are not novel compared to existing methods.

  • Distribution alignment does not necessarily guarantee that poisoned samples receive lower weights. The paper lacks direct analysis of the learned weight distributions for clean and poisoned samples.

  • The experiments only consider 5% and 10% trusted clean reference data, leaving performance under smaller or less representative reference sets unexplored.

  • Lack of adaptive analysis, where the attacker knows the design of the proposed method.

Questions for authors

Thank you for submitting the interesting work to ICLR 2027, but I still have a major concern about the robustness of the proposed method against attacks without latent separability.

The proposed method implicitly assumes that poisoned samples induce noticeable distribution shifts in the feature space, allowing KMM to assign them lower weights. However, this assumption may not always hold, as previous studies have shown that poisoned and clean samples can have highly overlapping feature distributions, such as Adap-Blend[1] and Grond[2]. In such cases, poisoned samples may still receive high weights, potentially undermining the effectiveness of DAR. I encourage the authors to evaluate DAR against these attacks and analyze whether the learned weights can effectively distinguish clean and poisoned samples when their feature distributions overlap.

[1]Revisiting the Assumption of Latent Separability for Backdoor Defenses, ICLR 2023

[2]Towards Backdoor Stealthiness in Model Parameter Space, CCS 2025

Reviewer_9BpWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Distribution-Aligned Reweighting (DAR) to defend against backdoor attacks during training. The main idea is to treat backdoor poisoning as a distribution shift between poisoned training data and clean data. The method uses a trusted clean reference set and class-wise Kernel Mean Matching to assign weights to training samples, so that their feature distributions are closer to the clean distribution. These weights are then used to guide model training without changing the model architecture or loss function. The authors evaluate DAR on three datasets against seven backdoor attacks with different poisoning ratios.

Critical strengths
  1. Instead of relying on differences in the learning dynamics of clean and poisoned samples, the paper explores distribution matching as an alternative way to reduce the influence of poisoned samples during training.
  2. DAR combines class-wise KMM with weighted sampling and does not require changes to the final model architecture or training objective. This makes the approach simple and potentially applicable to different classification models.
  3. The authors consider seven backdoor attacks, five poisoning ratios, and three datasets. The comparisons report clean accuracy, attack success rate, and robust accuracy, providing a more complete assessment than attack success rate alone.
Critical weaknesses
  1. The main method relies on existing KMM, followed by weight normalization, thresholding, and weighted sampling. The technical novelty beyond applying these existing techniques to backdoor defense is not sufficiently clear. Please clarify the contributions.
  2. The method assumes that poisoned samples cause feature-distribution shifts. However, distribution alignment does not necessarily assign lower weights to poisoned samples, especially when their features overlap with clean samples. An analysis of the learned weights for clean and poisoned samples would help support the proposed approach.
  3. DAR requires a trusted clean reference set. How can such a reference set be constructed in practice if no verified clean data are available? This assumption may limit the method's applicability.
  4. DAR combines several components, including class-wise distribution matching, label-shift adjustment, weight normalization, and adaptive cutoff. It is unclear how much each component contributes to the final performance. More detailed ablation studies would be helpful.
  5. The experiments mainly focus on all-to-one attacks with a fixed target class. It would be helpful to evaluate whether DAR also works against clean-label attacks, and more challenging attacks where poisoned samples have similar feature distributions to clean samples.
Questions for authors
  1. What is the main technical novelty of DAR beyond combining existing KMM, weight normalization, thresholding, and weighted sampling techniques? Could the authors clarify the key contributions?
  2. How does DAR ensure that poisoned samples receive lower weights, especially when their feature distributions overlap with those of clean samples? Could the authors provide an analysis of the learned weights for clean and poisoned samples?
  3. How can a trusted clean reference set be constructed in practice when verified clean data are unavailable? How would DAR perform without such a reference set?
  4. How much does each component, including class-wise distribution matching, label-shift adjustment, weight normalization, and adaptive cutoff, contribute to the final performance? Could the authors provide more detailed ablation studies?
  5. How well does DAR generalize to clean-label attacks or more challenging attacks where poisoned samples have feature distributions similar to clean samples? Could the authors include additional experiments to evaluate these settings?
Reviewer_1fuxOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Distribution-Aligned Reweighting (DAR), an in-training backdoor defense that treats data poisoning as an adversarial shift of the training distribution. Given a trusted clean reference set, DAR extracts feature representations and applies class-wise Kernel Mean Matching (KMM) to estimate sample weights. The weights are subsequently adjusted according to class proportions, normalized, and filtered using an adaptive cutoff before being incorporated into standard model training through weighted sampling.

Experiments on CIFAR-10, GTSRB, and ImageNet-12 cover seven backdoor attacks and five poisoning ratios. The proposed method achieves low average attack success rates while maintaining competitive clean accuracy and improving robust accuracy over the evaluated defenses. Additional experiments investigate reference-set size, model architectures, computational efficiency, and an adaptive attack.

Critical strengths
  1. The paper presents a straightforward distribution-matching perspective on backdoor defense, avoiding explicit assumptions about the learning dynamics of poisoned samples.

  2. The experiments cover multiple datasets, poisoning ratios, and attack types. DAR achieves particularly strong average ASR and RA results compared with ABL, DBD, and ESTI.

  3. The proposed approach does not require modifications to the final model architecture or training objective, and its end-to-end runtime is competitive with several existing defenses.

Critical weaknesses
  1. The methodological novelty is limited. The core components, including KMM, clean-reference-guided sample reweighting, and weighted sampling, are well-established techniques. The paper does not sufficiently distinguish its contributions from existing importance-weighting approaches, such as Gretton et al. [1], Ren et al. [2], and Shu et al. [3]. The current contribution appears primarily to be an application and combination of existing methods.

  2. The evaluation does not adequately isolate the benefit of the proposed reweighting strategy from the availability of trusted clean data. The main experiments use a clean reference set containing 10% of the original training data, which is a substantial resource. Important comparisons are missing, including training or fine-tuning directly with the same clean reference set, simple feature-similarity-based filtering, and existing clean-reference-guided reweighting approaches.

  3. Several methodological choices lack sufficient justification. Equation (2) omits the approximate weight-sum constraint commonly used in standard KMM. Moreover, the adaptive cutoff in Equation (5) is centered at one despite the preceding normalization enforcing a global rather than class-wise mean of one. The paper does not clearly establish the necessity of these choices or the separate contributions of KMM, label-prior adjustment, and adaptive cutoff.

  4. The robustness claims are broader than the experimental evidence supports. The evaluation focuses on all-to-one attacks without systematic coverage of clean-label or all-to-all settings. Additional attacks, such as WaNet [4] and Input-Aware Dynamic Backdoor [5], would strengthen the evaluation. Moreover, DAR reduces average clean accuracy on ImageNet-12 from 90.63% to 78.92%, and robust accuracy remains low for certain attacks. For example, ImageNet-12 under LowFreq at a 5% poisoning ratio achieves only 28.36% RA despite a 0.55% ASR. The adaptive attack evaluation is also restricted to one feature-alignment-based strategy rather than directly targeting the complete reweighting pipeline.

  5. The presentation and scalability evaluation need improvement. Tables 1–4 are excessively dense, especially Table 4, making the experimental results difficult to read and compare. The largest dataset is a 12-class ImageNet subset, leaving scalability to larger classification tasks unclear. The supplied manuscript also references Appendices B–D for important implementation and ablation details, but these appendices are not included in the provided PDF.

Ref:

[1] Covariate Shift by Kernel Mean Matching. In Dataset Shift in Machine Learning.

[2] Learning to Reweight Examples for Robust Deep Learning. ICML.

[3] Meta-Weight-Net: Learning an Explicit Mapping For Sample Weighting. NeurIPS.

[4] WaNet — Imperceptible Warping-based Backdoor Attack. ICLR.

[5] Input-Aware Dynamic Backdoor Attack. NeurIPS.

Questions for authors
  1. Can the authors compare DAR with simple baselines using exactly the same trusted clean reference data, including clean-only training, clean-data fine-tuning, feature-similarity-based filtering, and representative sample-reweighting methods?

  2. What are the individual contributions of class-wise KMM, label-prior correction, and adaptive cutoff? How does performance change when using the standard KMM weight-sum constraint or a cutoff centered at the class-wise mean? An analysis of the learned weight distributions for clean and poisoned samples would also be helpful.

  3. Can the authors extend the evaluation to clean-label, all-to-all, and additional adaptive attacks? In particular, can an attacker optimize poisoned samples to retain large final sampling weights after the complete KMM and cutoff procedure?

  4. Can the authors explain the substantial clean-accuracy degradation on ImageNet-12 and the low robust accuracy under certain attacks, and provide experiments with smaller trusted reference sets and larger-scale datasets?

Reviewer_PdH6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DAR (Distribution-Aligned Reweighting), an in-training backdoor defense that frames poisoning as an adversarial shift of the training distribution away from the underlying clean distribution. Given a small trusted clean reference set, DAR performs class-wise feature distribution alignment using kernel mean matching (KMM) to assign weights to training samples, applies an adaptive class-specific cutoff, and uses the resulting weights to construct a sampling distribution for standard model training. The method requires no prior knowledge of the attack and does not modify the model architecture or training objective. Experiments on CIFAR-10, GTSRB, and ImageNet-12 across seven backdoor attacks and five poisoning ratios demonstrate that DAR achieves low attack success rates while maintaining competitive clean and robust accuracy, compared to ABL, DBD, and ESTI. An adaptive attack experiment is also included.

Critical strengths

The primary strength is the clean conceptual framing of backdoor poisoning as an adversarial distributional shift, which motivates a principled reweighting strategy via kernel mean matching rather than relying on heuristic loss-based separation signals.

The method is broadly applicable: it requires no modification to model architecture or training objective, and the weight computation is decoupled from the downstream trainer, offering practical flexibility across different backdoor defense scenarios.

Critical weaknesses
  1. My fundamental concern is the reliance on a trusted clean reference set, which is a strong and often unrealistic assumption in practical deployment scenarios. The paper treats the reference set as a given prerequisite without critically examining the implications of its size or potential contamination. In real-world settings where a defender may have access to only a handful of clean samples—or where the reference set itself could be partially poisoned by an adversary with write access to data pipelines—the distributional alignment signal may become unreliable or even counterproductive, amplifying the very shifts DAR aims to correct.

  2. In addition to the reference set assumption, the baseline comparisons are notably incomplete. The paper only compares against ABL, DBD, and ESTI, omitting several important in-training and post-training defenses that have been proposed more recently. Notably absent are methods such as Neural Cleanse (Tran et al., 2019), which employs spectral analysis for backdoor detection, and more recent approaches that leverage representation learning or contrastive objectives for defense. Without comparison to these methods, the claimed superiority of DAR remains incompletely substantiated, and it is unclear whether the distributional alignment perspective offers genuine advantages over alternative paradigms.

  3. Furthermore, the computational overhead of DAR warrants closer scrutiny. As shown in Table 5, the total wall-clock time for DAR on ImageNet-12 is 25,140 seconds, which is comparable to ESTI (26,207 seconds) but significantly higher than ABL (4,529 seconds). The feature extractor training stage alone accounts for 12,576 seconds, and while the paper notes that a randomly initialized network yields comparable performance, this still requires a full training pass over the poisoned data to obtain features—defeating part of the efficiency motivation. A more thorough ablation on whether the feature extractor training step is strictly necessary, or whether a cheaper approximation suffices, would strengthen the practical argument for DAR.

Questions for authors
  1. How robust is DAR's performance to contamination of the trusted clean reference set? If even a small fraction of the reference samples are poisoned, how does the distributional alignment objective behave, and is there a threshold of reference set corruption beyond which defense performance degrades sharply?

  2. The paper compares against ABL, DBD, and ESTI but omits several other notable backdoor defense methods. Could you clarify the design space coverage of your baselines and discuss whether the distributional alignment perspective could be combined with or outperformed by methods operating at different stages of the pipeline (e.g., pre-training data cleaning or post-training pruning)?

  3. Table 5 reports that the feature extractor training stage dominates DAR's computational cost on ImageNet-12. Since the paper also shows that a randomly initialized network yields comparable defense performance, could the feature extraction step be replaced by a fixed, pretrained feature extractor (e.g., from a self-supervised model) to reduce overhead without sacrificing defense quality?

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

Related papers

Discussion (0)

Sign in to comment.