acceptodds
Withdrawn from ICLR 2027

DualSpec: Exposing Disguised Federated Backdoor via Cross-Client Structure and Trajectory Residuals

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated learning (FL) is vulnerable to backdoor attacks that manipulate client updates to implant backdoor features into the global model. Existing server-side defenses often rely on observable discrepancies in update magnitude, coordinate-wise statistics, or on comparisons against a fixed reference model in isolation, which may become unreliable against disguised malicious updates or adaptive trigger attacks. Other approaches rely on auxiliary or trusted datasets, introducing additional data requirements. We propose , an FL backdoor defense that combines two complementary detection signals, without requiring auxiliary data. The branch applies singular value decomposition (SVD) to per-layer normalized client updates to obtain a shared spectral space across clients, where suspicious groups are identified based on their directional cohesion and isolation. The branch instead applies SVD to the global model's own recent update to construct a per-layer reference subspace, and compares each client's update against it using residual energy to identify updates that deviate from the recent trajectory. Both branches derive from spectral decomposition of model updates, which exposes their underlying structure beneath surface-level disguise. Finally, malicious updates detected from both branches are combined before aggregation to capture backdoor behavior at both the group and the individual level. Experiments across different datasets and IID/non-IID settings show that DualSpec consistently suppresses a diverse set of state-of-the-art backdoor attacks while preserving main task accuracy, and generalizes beyond convolutional architectures.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Backdoor Defense, Singular Value Decomposition

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_R5LpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DualSpec, a server-side detection framework for federated learning backdoor defense. By exploiting the structural characteristics of malicious client updates, the authors design two complementary branches: Cross-client Spectral Detection, which identifies coordinated attack groups through spectral embedding and clustering, and Trajectory-residual Detection, which detects abnormal updates by measuring the deviation between client updates and the global trajectory subspace. DualSpec filters suspicious client updates before aggregation and achieves lower attack success rates while maintaining high main-task accuracy under various backdoor attacks and IID/non-IID settings.

Critical strengths
  1. Clear motivation. The paper starts from the structural characteristics of client updates and provides a new perspective for federated learning backdoor defense by analyzing cross-client spectral structures and the deviation of client updates from the global optimization trajectory.

  2. Comprehensive experiments. The paper conducts experiments under various backdoor attacks, IID/non-IID data distributions, and different model architectures, and compares the proposed method with multiple defense approaches.

Critical weaknesses
  1. The paper evaluates adaptive attacks such as A3FL and BATMAN; however, these attacks are not specifically designed to target DualSpec’s detection mechanisms. Since DualSpec fundamentally relies on cross-client spectral structure and trajectory-residual signals, a more targeted attacker could explicitly constrain malicious updates to simultaneously reduce cross-client cohesion and trajectory residuals, thereby evading detection.

  2. The paper reports the malicious elimination rate (MER), but does not report the proportion of benign client updates that are incorrectly removed. Since DualSpec directly excludes flagged client updates before aggregation, such false removals may not only affect detection reliability, but also reduce the contribution of benign clients to global model training. Although the paper reports high main-task accuracy (MA), this metric cannot directly reveal whether benign clients are frequently filtered out. The authors are encouraged to report the benign-client false positive/elimination rate and further analyze the impact of such false removals on benign clients’ training contributions and the final model performance.

  3. The authors use BadNets, Chameleon, and Neurotoxin to observe two types of phenomena and then use these observations as the basis for the design of the DualSpec framework. However, the Motivation section does not demonstrate that other attacks, such as DBA, BATMAN, and A3FL, necessarily follow the same two types of structures.

Questions for authors
  1. The authors should further consider defense-aware adaptive attacks and evaluate whether an attacker with knowledge of DualSpec can still maintain a high ASR while bypassing both detection branches.

  2. The authors should report the benign-client false positive/elimination rate and further analyze the impact of such false removals on benign clients’ training contributions and the final model performance.

  3. The framework is motivated by observations from BadNets, Chameleon, and Neurotoxin. How do the authors justify that the identified spectral cohesion and trajectory deviation patterns are general properties of federated backdoor attacks rather than attack-specific behaviors? Do attacks such as DBA, BATMAN, and A3FL exhibit the same underlying structures?

Reviewer_Wzv8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces DualSpec, a server-side defense against federated backdoor attacks that combines cross-client clustering with per-layer residual analysis relative to the previous global update. The method excludes clients identified by either component without requiring auxiliary data. Against A3FL on CIFAR-10, DualSpec achieves 2.22% ASR with 87.65% clean accuracy, compared with 100% ASR for AlignIns and 28.41% for DnC. On CIFAR-100, it achieves 5.91% ASR, compared with 16.29% for AlignIns.

Critical strengths
  1. The two branches have a clear empirical motivation, supported by an informative ablation. In Figure 3, the cross-client branch suppresses BadNets and Chameleon but leaves Neurotoxin at approximately 67.5% ASR; the trajectory branch suppresses Neurotoxin but leaves the other two attacks largely effective. Combining them reduces ASR to below 1% in these settings.

  2. The main comparison includes relevant recent attacks and defenses, including BATMAN (CVPR 2026), AlignIns (CVPR 2025), and FedDLAD (IJCAI 2025). The A3FL result is particularly substantial: on CIFAR10, DualSpec reports 2.22% ASR versus 100% for AlignIns and 28.41% for DnC, while maintaining 87.65% clean accuracy. On CIFAR100, its A3FL ASR is 5.91% versus 16.29% for AlignIns.

  3. The evaluation covers delayed attacks, different malicious-client and poisoning ratios, IID and heterogeneous partitions, blended triggers, and additional modalities. The Gram-matrix implementation is appropriate for a small number of clients and high-dimensional updates. The paper also reports a failure regime: Neurotoxin becomes substantially more effective at a 40% malicious-client ratio.

Critical weaknesses
  1. The reported detection metric does not measure erroneous removal of honest clients. MER measures malicious-client removal, while the clustering branch selects the highest-scoring candidate cluster without an explicit abstention threshold. Consequently, an honest cluster could be excluded even in an attack-free round. Report benign-client rejection rates and clean accuracy during attack-free training and before delayed attacks, particularly under non-IID partitions.

  2. The evaluation does not test an attacker explicitly targeting DualSpec's two detection signals. LGA, which the paper cites but does not evaluate, aligns malicious updates layer-wise with the previous global update and therefore directly challenges the trajectory-based detector. Moreover, Appendix A.4 states that Expansion does not handle a second independent malicious group. Evaluate LGA and multiple malicious groups, together with an adaptive attack that jointly minimizes clustering and residual scores.

  3. The experiments do not isolate the benefit of the spectral representation over simpler clustering. FLAME already uses HDBSCAN, whereas DualSpec also changes normalization, rank truncation, and cluster selection. Figure 3 establishes branch complementarity only for BadNets, Chameleon, and Neurotoxin. Hold the clustering and selection rules fixed, compare normalized updates with their rank-truncated representations, and evaluate the residual branch against a simpler layer-wise alignment score.

  4. The reported comparisons lack uncertainty estimates and a clear tuning protocol. Several improvements over AlignIns outside A3FL are small, but the tables provide neither independent run counts nor confidence intervals. The dataset-dependent shape weight also lacks a stated held-out selection procedure. Report repeated-run results, specify how hyperparameters are selected without using evaluation attacks, and measure server runtime and peak memory.

Questions for authors
  1. What are the benign-client false-positive rate, detection precision, and clean-accuracy cost in attack-free training and before delayed attacks? Can the cross-client branch abstain when all candidate clusters are benign, and how is that decision calibrated?

  2. Does DualSpec remain effective against LGA and an attacker explicitly seeking to evade both branches? How does it behave when attackers form multiple independent groups or influence the previous global reference over time?

  3. Starting from a common HDBSCAN-based pipeline related to FLAME, what gains come separately from layer normalization, spectral truncation, the proposed cluster-selection rules, and the trajectory branch? Can the authors report rank sensitivity and compare the residual detector with a simpler layer-wise alignment score, while measuring both malicious and benign rejection?

  4. Can the main results be repeated with uncertainty estimates, a held-out tuning protocol, and measured runtime/memory, with the missing implementation choices specified?

Reviewer_T5EiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DualSpec, a server-side defense against backdoor attacks in federated learning that does not require auxiliary or trusted server data. DualSpec consists of two complementary detection branches. The Model-wise Cross-client branch applies per-layer normalization and spectral embedding to client updates, followed by HDBSCAN clustering and cohesion–isolation scoring to identify coordinated malicious clients. The Layer-wise Trajectory-residual branch projects each client’s per-layer update onto a reference subspace derived from the most recent global update and flags clients with unusually high residual energy. The flagged clients from both branches are combined and excluded before aggregation.

Critical strengths
  • The paper addresses an important problem, and the two-branch design is intuitively motivated.

  • The evaluation covers a broad range of attacks, data distributions, and model architectures.

  • The paper is well-organized and easy to follow.

Critical weaknesses
  • In Algorithm 1, whenever HDBSCAN returns at least one candidate cluster, the method selects the highest-scoring cluster through argmax and marks it as suspicious. There is no explicit criterion for deciding that none of the candidate clusters is malicious or that no client should be excluded in the current round. Consequently, even in a completely benign FL system, the method may select and remove a benign cluster. Taking the union of the current-round detection, two-round detection, and trajectory-residual detection may further increase the number of excluded benign clients.

  • The paper primarily reports malicious elimination rate (MER), but MER alone does not characterize detection quality. The results of False-positive rate and benign-update exclusion rate; Detection precision and recall; The average number or fraction of clients retained per round; Experiments in a completely attack-free setting are highly preferred.

  • Figure 3 demonstrates complementarity between the two branches, and Table 11 examines the refinement components. However, the paper does not yet provide direct evidence that the proposed spectral embedding itself is necessary. Whether the improvements come from the multi-component spectral representation, per-layer normalization, cluster scoring, or the combination of several heuristic mechanisms should be further clarified.

  • The reference subspace in Eqs. (9)–(11) is derived from a single previous global-update matrix, rather than from a sequence of updates across multiple rounds. The authors should clarify the meaning of “trajectory subspace” and justify why the left-singular subspace of one global update adequately represents the recent optimization trajectory.

-The main text uses MA, whereas Figure 2 and Table 11 use CA. The notation should be unified.

-Empirical runtime, memory consumption, and scalability with the number of participating clients would complement the asymptotic complexity analysis.

Questions for authors

See weakness

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

Related papers

Discussion (0)

Sign in to comment.