acceptodds
Under review as a conference paper at ICLR 2027

COEX-FL: Non-Colluding Backdoor Coexistence in Federated Learning via Semantic–Mode Coding

Abstract

Many federated backdoor attacks assume that malicious clients can collude by sharing target labels, trigger parameters, training schedules, or model updates. Such coordination introduces communication and synchronization requirements and does not capture settings in which clients are compromised independently. Without coordination, multiple backdoors may induce conflicting update directions, causing some target behaviors to be attenuated or forgotten during aggregation. We propose COEX-FL, a coordination-free framework for multi-target backdoor coexistence in federated learning. COEX-FL factorizes each frequency-domain trigger into a target-semantic component and a mode-specific realization component. Orthogonal semantic components distinguish target behaviors within a shared carrier space, while multiple mode realizations provide alternative physical paths for the same target. Each malicious client constructs its trigger bank from public coding parameters and its own target label, and optimizes over the modes locally without exchanging private attack information with other clients. We characterize the resulting code and DCT-space geometry and derive a conditional linearized analysis of cross-attacker update interference. Across the evaluated datasets and federated learning protocols, COEX-FL improves the balance and retention of multiple backdoors, achieving high minimum attack success rates and low cross-attacker interference.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
federated learning, backdoor attack, trustworthy federated, ai security

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_H3DNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes COEX-FL, a federated attack framework for multi-target, multi-shape backdoors, where the malicious clients do not communicate during weight updates. Each attacker independently builds frequency-domain triggers from a shared Hadamard-based coding scheme.

Critical strengths

The backdoor triggers has training-free closed form construction, derived from the shared parameters and its own labels, requiring no runtime communication.

Critical weaknesses
  1. Authors claim COEX-FL to be "non-colluding", however, all adversaries must share the same attack parameters and strategies. It is only that no information is exchanged during training. Authors should reposition and reframe the storyline. If Authors want to stick with the "non-colluding" setting, additional experiments should be added when other attackers can adopt different attack schemes.

  2. The writing is very difficult to follow, especially in sec.3 methodology. Many notations are never explained. Eg: "mode" is a very important concept that describes the attack setting, and appears in the title, but authors never defined this term throughout the paper.

Questions for authors

How should one pick the hyperparameters like ρ,λ,β\rho, \lambda, \beta? The writing suggests that these are controllers/gates of the attack strength, but no systematic ablations are provided for hyperparameter sensitivity.

Reviewer_1jz9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies backdoor attacks in federated learning (FL) through the lens of cross-client interference. Many existing attacks assume that malicious clients cooperate within a single attacker group. However, this assumption does not capture settings with independently operating attackers. Different groups of attackers can interfere with each other, potentially degrading backdoor effectiveness. In light of this, the authors introduce COEX-FL, which factorizes triggers into target-conditioned and physical-realization components to reduce interference. Experiments across multiple datasets show that COEX-FL improves multi-backdoor effectiveness compared with existing methods, such as Mirage.

Critical strengths
  • The problem of interference among non-colluding attackers is an interesting angle for studying backdoor attacks in FL.
  • The idea of factorizing the target signal and its physical realization is sound and interesting.
Critical weaknesses
  • The core idea of this paper is about non-colluding attackers, but the experiments focus on attackers using the same attack method. This could be problematic because COEX-FL assumes that attackers adopt a shared public construction and configuration. The justification for this assumption remain unclear.
  • The interference-induced performance degradation is not clearly demonstrated. While the authors mention that the presence of different groups of attackers can reduce attack performance, this claim is not well supported by empirical evidence. In particular, it is understandable that this might occur when different groups use the same trigger but assign it to different targets. However, trigger collisions are very rare in practice. Any backdoor attack with an optimizable trigger pattern seems capable of addressing this issue.
  • While the authors provide a link to the source code, the paper does not clearly explain how to reproduce the comparisons or how the baselines are adapted to this setting. This is particularly important for assessing experimental fairness. Baselines such as DBA were originally designed around coordinated attackers. It is crucial to clarify how these methods are adapted and whether their configurations and tuning budgets are comparable to those of COEX-FL.
  • While the experimental results are promising, the analysis of defense robustness is limited. The authors offer a qualitative explanation, but more defense-specific evidence is needed to explain why COEX-FL remains effective against the evaluated defenses.
Questions for authors

Please refer to the weaknesses above.

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

This paper studies federated backdoor attacks in which multiple attackers pursue their own targets without exchanging attack information during training. The aim is to reduce interference so that several backdoors can coexist in one global model. COEX-FL constructs frequency-domain triggers from a shared public configuration and each attacker's target label. A target-specific code separates target identities in code space, while multiple modes provide alternative realizations of the same backdoor. The evaluation covers continuous and single-round injection, defenses, interference between attackers, and retention after attacks stop. The reported results suggest improvements in both average attack success and the weakest attacker's performance.

Critical strengths
  • The paper studies how multiple backdoors interfere with each other in federated learning, which is a useful problem.

  • The trigger construction is simple and clearly described.

  • The experiments report individual attackers' success rates and test how long the backdoors remain after attacks stop.

Critical weaknesses
  1. I am not convinced by the non-colluding claim. The paper defines non-collusion through the absence of runtime information exchange, but assumes that all attackers adopt the same public configuration before training. In my view, this already builds prior coordination into the attack design. The method's compatibility comes from that shared setup, which makes the connection to independently compromised clients unconvincing. This is my main concern with the proposed approach.

  2. Generality across datasets is not established. The code length and carrier allocation depend on the label space, while frequency placement must match the input. It is unclear which settings were kept fixed across datasets and which were adjusted. The paper does not test whether parameter choices can be reused on another dataset without retuning, and provides no systematic sensitivity analysis for perturbation strength, the semantic/mode balance, mode count, or frequency placement.

  3. Some experimental results need stronger support. Neurotoxin achieves only 1.23% average ASR on CIFAR-10 in Table 1, while FCBA has 18.62% clean accuracy on GTSRB in Table 2. COEX-FL's higher clean accuracy on Tiny-ImageNet also needs explanation. Clearer baseline adaptation and tuning details are needed to interpret these gaps.

  4. The analysis leaves important results unexplained. The ablation refers to "target-indexed carrier separation," whereas Section 3.3 states that all targets share the semantic carrier. It is unclear what this ablation removes. The Full CIFAR-10 average ASR also differs between Tables 1 and 4 (99.80% versus 95.38%) without a clear explanation of the settings.

Questions for authors
  1. Which settings are fixed across datasets, and which are retuned?
  2. How were the baselines adapted and tuned, and how many runs support Tables 1-4?
  3. What does the carrier-separation ablation remove, and why do the Full results differ between Tables 1 and 4?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.