acceptodds
Withdrawn from ICLR 2027

Anchor in Evidence: Distribution-Preserving Backdoor Attacks in Federated Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated learning (FL) enables collaborative model training without sharing local data, but its decentralized nature also makes it vulnerable to backdoor attacks. Among these attacks, in-distribution (ID) backdoor attacks are particularly attractive because they preserve distribution consistency and are difficult to detect. However, we observe that existing ID backdoor attacks become highly unstable under heterogeneous client distributions, where attack success can abruptly deteriorate during federated training. We identify this previously overlooked phenomenon as *Attack Collapse* and show that it arises from the continuously evolving feature space in FL, which progressively misaligns injected backdoor representations with the target-class distribution. Motivated by this observation, we propose IDEA, a distribution-preserving backdoor attack that maintains stable target alignment throughout federated training. IDEA first leverages evidential deep learning to identify reliable target-class samples and construct robust target regions. It then jointly performs region-aware alignment and evidential geometric refinement to preserve both target affinity and local semantic structure, enabling poisoned samples to remain compatible with the evolving feature space. Extensive experiments on multiple benchmark datasets demonstrate that IDEA consistently mitigates Attack Collapse and significantly improves attack efficiency and stability over state-of-the-art backdoor attacks, while maintaining competitive clean-task performance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Backdoor Attacks, In-Distribution Poisoning, Attack Stability, Evidential Deep Learning

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

This paper focuses on backdoor attacks in federated learning. The authors observe that client heterogeneity affects attack stability and refer to this phenomenon as Attack Collapse. They further attribute this instability to the misalignment between poisoned features and the target-class distribution. To address this issue, the authors construct stable anchors using EDL, guide poisoned samples toward feature regions compatible with the target class, and meanwhile maintain appropriate semantic separation from non-target classes. Extensive experiments demonstrate the effectiveness of the proposed method.

Critical strengths

The authors observe that increasing client heterogeneity can lead to significant ASR fluctuations, which they characterize as Attack Collapse. I find this phenomenon interesting and believe it has a certain degree of novelty and practical significance.

The experimental evaluation is relatively comprehensive, the overall results are strong, and the authors also provide the corresponding code.

Critical weaknesses
  1. I find the phenomenon identified by the authors interesting. However, I believe the main issue is that the current analysis does not yet establish a complete and sufficiently substantiated chain of reasoning. Why does stronger client heterogeneity lead to more severe Attack Collapse? Figure 1 shows that ASR becomes increasingly unstable as heterogeneity increases, but it remains unclear whether this instability is indeed caused by feature misalignment. More convincing evidence is needed to support this claim.

  2. The authors introduce EDL to enable more reliable anchor selection, and the motivation is reasonable. However, the paper does not provide sufficient quantitative or qualitative experiments to demonstrate the necessity of this module.

  3. The paper lacks direct quantitative validation of its “distribution-preserving” property. The current Acc, ASR, AQ, and t-SNE results mainly demonstrate attack effectiveness and stability, but do not sufficiently verify that poisoned samples indeed remain within the target data distribution.

  4. Lemma 2 and Theorem 1 require anchors to satisfy threshold conditions on concentration and conflict, respectively. However, the algorithm selects anchors solely by ranking the product of the two, i.e., ADS. Therefore, it remains unclear whether this selection strategy can guarantee that the conditions required by the theoretical analysis are satisfied.

Questions for authors
  1. Could the authors provide stronger evidence to support the claim that more severe Attack Collapse under higher client heterogeneity is mainly caused by feature misalignment, rather than simply showing the corresponding ASR fluctuations?

  2. Could the authors provide additional quantitative or qualitative evidence to demonstrate the necessity of EDL for reliable anchor selection, for example by comparing it with simpler confidence-based or probability-based anchor selection strategies?

  3. Could the authors provide direct quantitative evaluation of the “distribution-preserving” property to verify that poisoned samples indeed remain close to the target-class distribution?

  4. Could the authors clarify how the ADS-based ranking used in the algorithm ensures that the selected anchors satisfy the separate concentration and conflict threshold conditions required by Lemma 2 and Theorem 1?

Reviewer_z3SbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the instability of in-distribution (ID) backdoor attacks in heterogeneous federated learning, identifying a phenomenon termed Attack Collapse, where attack success rates can drop sharply during training. To address this issue, the authors propose IDEA, which selects reliable target-class anchors using evidential deep learning, constructs target regions through clustering, and optimizes poisoned samples using region-aware alignment and evidential geometric refinement. Experiments across multiple datasets, model architectures, and defense mechanisms demonstrate improvements in attack success rate and stability over existing baselines.

Critical strengths

Meaningful problem and interesting observations. The instability of backdoor attacks under heterogeneous federated training is a relevant problem. The observed abrupt fluctuations in ASR provide an interesting perspective beyond conventional evaluations focused primarily on final attack success.

Encouraging empirical results. The paper provides extensive experiments across different datasets, defenses, and model architectures. The improvements over strong baselines, particularly Mirage, are promising, and the ablation studies provide initial evidence for the effectiveness of the proposed components.

Critical weaknesses

W1. Insufficient differentiation from closely related work.

The paper's central motivation and optimization strategy substantially overlap with prior studies on persistent and feature-aligned backdoor attacks, including Chameleon, A3FL, Mirage, and Parasite. In particular, target-feature attraction, separation from competing representations, and maintaining backdoor effectiveness under evolving federated models have been investigated previously. Although IDEA introduces evidential anchor selection and region-based optimization, the paper does not sufficiently establish what fundamentally new capability these components provide. More detailed theoretical and experimental comparisons are needed to clarify the novelty and necessity of the proposed design.

W2. Mathematical inconsistencies in the core anchor selection mechanism and discrepancies between theory and implementation.

The Anchor Discovery Score (ADS) is defined as the product of prediction concentration and opinion conflict, with smaller values preferred. However, the concentration term can become negative when the maximum predictive probability exceeds 0.5. Consequently, minimizing ADS may favor samples with greater opinion conflict, contradicting the intended selection of reliable, low-conflict anchors.

Additionally, Eq. (11) and Algorithm 1 use inconsistent definitions of the anchor set size. The theoretical optimization analysis also considers projected sign-gradient updates, whereas the implementation details describe Adam-based trigger optimization. These inconsistencies require clarification, particularly because anchor reliability and optimization stability are central to the proposed method.

W3. Insufficient causal evidence for the proposed explanation of Attack Collapse.

The authors attribute Attack Collapse primarily to feature-space evolution and the resulting misalignment between poisoned samples and target-class representations. However, the current evidence relies heavily on ASR trajectories and feature-space visualizations, which establish correlation rather than causation. Alternative explanations, such as variations in client participation, aggregation dynamics, or defense filtering, have not been sufficiently excluded. Moreover, the existing ablation studies demonstrate the usefulness of individual components but do not directly establish that IDEA improves stability by mitigating the proposed feature-misalignment mechanism.

Questions for authors

Novelty and related work: Could the authors clarify the fundamental differences between IDEA and existing persistent backdoor attacks, particularly Mirage, Chameleon, and Parasite? A more explicit theoretical discussion and targeted experimental comparisons with simpler prototype-alignment or contrastive-learning alternatives would help establish the unique contribution of evidential anchor selection and geometric refinement.

Mathematical correctness and implementation: How do the authors reconcile the potentially negative prediction concentration term with the intended ADS ranking criterion? Could the authors also clarify the discrepancies between Eq. (11), Algorithm 1, and the optimizer used in the theoretical analysis versus the actual implementation?

Causal attribution: Can the authors provide more direct evidence that feature-space misalignment is the primary cause of Attack Collapse? For example, controlled experiments that jointly measure feature drift, target/non-target distances, ASR fluctuations, and client participation or defense outcomes would help distinguish the proposed explanation from alternative causes.

I am currently leaning toward rejection, but I consider the work potentially improvable. I would be willing to reconsider my assessment and raise my scores if the rebuttal convincingly solve my concerns.

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

This paper studies in-distribution backdoor attacks in heterogeneous federated learning. It identifies Attack Collapse, in which ID backdoors drift away from the evolving target feature region, degrading ASR across rounds. The paper proposes IDEA, which uses evidential deep learning to select reliable target-class anchors, clusters them into region centers, and optimizes triggers with region-aware attraction and evidential geometric refinement to make poisoning effective and stable across rounds.

Critical strengths
  1. Great problem and sensible method. The problem of ID backdoors in heterogeneous FL, called Attack Collapse in this paper, is real but underdeveloped. The paper investigates this problem and offers valuable insights. The general idea of finding the closest stable region and other refinements is straightforward and reasonable, and the concrete method also seems generally plausible.

  2. Thorough evaluation. The experimental coverage is quite complete.

Critical weaknesses
  1. Important differences/gaps between the claims, design, and implementation that impair the soundness of this paper. a) An important concept used by this paper, EDL, is not correctly implemented or described. The global model used in the experiment, ResNet-18, is not an EDL model. According to the code, the paper does not attempt to convert it. Instead, they train it in the usual way and calculate softplus(logits)+1 as evidence when training the trigger. In my opinion, the “evidence” obtained this way has nothing to do with EDL; it's just a simple transformation of the logits. This mismatch, therefore, invalidates the theoretical assumptions and the seemingly reasonable steps that follow. Even if the paper implemented it correctly, this method is still not general, since most models are not EDL models, and the paper should mention this. b) The evaluation protocol is inconsistent between the design and implementation. The appendix states that there are 100 clients, 20 of which are malicious, and that 10 are selected each round, including 2 malicious ones. However, the code and log suggest that only 2 malicious clients participate in all rounds. This gives malicious clients an unfair advantage and makes the result unreliable. This is not the only inconsistency, and I suggest a thorough implementation check. c) The cosine term in Eq. (1) is questionable. First, this term is also implemented incorrectly. In the paper, it is cosine similarity between features, but in the code, it is between logits. Second, even if the implementation were correct, the purpose of this term is still unclear. d) The design of “distribution mapping” is too weak to support the claim. The paper claims that the “distribution mapping” can align triggered samples with the target distribution. But this design is naive and resembles an oversimplified version of an adversarial adaptation method. I cannot see how this design fulfills the claim without extra explanation or ablation experiments.

  2. The theoretical explanation stands alone and does not give any insights. Most of the theorems are just simple derivations from the assumptions. Experimental results support none of these assumptions, and some even contradict the method or other informal assumptions in this paper.

Questions for authors
  1. What is the purpose of the cosine term in Eq. (1)?
  2. Do the results still support your claims after fixing all the inconsistencies?
  3. Is “distribution mapping” actually effective? Please provide results to support this.

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

Related papers

Discussion (0)

Sign in to comment.