acceptodds
Withdrawn from ICLR 2027

OTPA: Federated Model Poisoning via Optimal Transport Alignment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated defenses often assess client behavior through update geometry, but parameter proximity need not imply benign functional influence. Motivated by neural permutation symmetry, Optimal Transport Poisoning Attack (OTPA) transports a stored poisoned source toward a clean, round-specific anchor before interpolation. Its entropic maps combine approximate feature alignment with mixing, without honest-client updates or explicit amplification. We specify transport propagation for convolutional, residual, vision-transformer, and BERT-style architectures. The primary vision benchmark compares nine baselines while distinguishing attacker-local from oracle information. OTPA reduces accuracy from 66.0% to 14.9% on Tiny-ImageNet and from 66.7% to 39.4% on CIFAR-10 while its updates remain close to the centre of the submitted population. On all three datasets it is Pareto-optimal among the evaluated attacker-local attacks, reducing the median update distance by 70% to 90% relative to matched unaligned interpolation at no more than accuracy points of difference. Further evaluations cover architecture transfer, language, heterogeneous data, and partial participation. OTPA yields the lowest defended accuracy under six robust aggregators, while history-based filtering exposes highly consistent malicious directions. These results establish transport-based source construction as a poisoning mechanism: proximity and directional consistency reveal different aspects of an update, so neither coordinate alignment nor low Euclidean distance establishes benign behavior.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Optimal Transport, Adversarial Machine Learning

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_R3PSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces OTPA, an untargeted model-poisoning attack for federated learning. Using local data and the broadcast model, it aligns a stored poisoned model with a freshly trained clean anchor through entropic optimal transport, then interpolates the two for submission. On CIFAR-10, OTPA reduces global accuracy from 66.69% to 39.41%, with a median update distance of 5.47 compared with 55.00 for unaligned interpolation. It outperforms the evaluated attack baselines under six robust aggregators, while FoolsGold and FLAME neutralize it in the reported settings.

Critical strengths
  1. The proposed attack produces a substantial impact-distance tradeoff. On CIFAR-10, OTPA reduces global accuracy from 66.69% to 39.41%, while its median update distance is 5.47 compared with 55.00 for matched unaligned interpolation. Across the three primary vision datasets, it reduces median distance by 70-90% relative to Linear while remaining within 2.5 accuracy points of its impact. Matching the source, anchor, interpolation weight, and schedule makes this a useful comparison of the complete transport operation.

  2. The evaluation includes actual robust aggregation mechanisms. OTPA leaves approximately 43-46% CIFAR-10 accuracy under Krum, Multi-Krum, trimmed mean, coordinatewise median, RFA, and FLTrust, lower than every evaluated baseline in those columns. These results provide more direct evidence of attack effectiveness than update distance alone.

  3. The construction is evaluated across multiple architectures, heterogeneous partitions, partial participation, and a language task. The paper distinguishes attacker-local baselines from those receiving honest-update information. It also reports effective countermeasures: FoolsGold and FLAME neutralize OTPA in the tested runs, and DnC substantially reduces its impact.

Critical weaknesses
  1. The incremental novelty over the closest prior attack is not sufficiently established. FL-MMR already uses optimal transport to align poisoned and benign models for federated poisoning, and the authors acknowledge it in Table 6. OTPA's persistent source, refreshed anchor, dense interpolation, and architecture coverage may provide useful differences, but the paper does not compare against FL-MMR experimentally or isolate the benefit of those differences. Existing transformer OT fusion also already addresses attention, normalization, and residual routing. The paper should identify which part of the attack is new and demonstrate its value relative to the closest predecessor under comparable information and resource assumptions.

  2. The results establish the value of the complete transport operation, but do not establish the attack benefit of learned feature correspondence beyond contraction and mixing. Appendix E.1 shows that a uniform map reduces source-anchor distance from 62.85 to 10.10, compared with 8.23 for solved OT. The authors acknowledge that contraction accounts for most of the reduction. However, these controls are evaluated only for a round-100 construction, without complete training trajectories or defended accuracy. This leaves open whether simpler contraction can reproduce the main attack results. It does not invalidate OTPA's observed effectiveness, but weakens the explanation of why its central design is needed.

  3. The statistical support for the main defense comparison is incomplete. Table 2 reports Student-t confidence intervals without specifying the replicate count and seed protocol for each condition. Table 4 provides no uncertainty or clean no-attack rows for its seven mechanisms, while the alignment ablation and explicit-filter evaluations are single runs. The reported margins under six aggregators are substantial, but repeated runs and defense-specific clean controls are needed to assess their stability and quantify the accuracy loss attributable to each attack.

  4. A smaller update distance does not necessarily imply greater stealthiness. Malicious updates influence the reference center, and comparisons with a separate clean run do not show whether OTPA resembles honest updates during the attack. The authors should compare malicious and honest updates within the same attacked run; the reported failures against FoolsGold and FLAME further limit claims of general defense evasion.

  5. The defense evaluation omits recent relevant methods, including FoundationFL (NDSS 2025), LASA (WACV 2025), and FedREDefense (ICML 2024). Comparisons under compatible assumptions would clarify whether OTPA’s effectiveness extends to these defenses, particularly given its limited success against DnC, FoolsGold, and FLAME.

Questions for authors
  1. What is the main technical advance over FL-MMR, and can the authors compare the two under matched attacker information, compromise rates, and resource budgets?

  2. Do uniform-map and contraction-matched randomized-map controls retain poisoning impact and evade the same aggregators over complete training runs? This would directly test whether learned correspondence is necessary for the reported advantage.

  3. How many independent runs support each confidence interval, and can the authors add uncertainty and defense-specific clean controls to Table 4?

  4. Does the geometric and defended-accuracy advantage persist when evaluated against honest-client geometry in the same run and at lower compromise rates?

  5. Can the authors evaluate representative recent defenses such as FoundationFL, LASA, and FedREDefense, with their required assumptions and defense-specific clean accuracy clearly reported? If a method is incompatible with the threat model or protocol, please explain the incompatibility.

Reviewer_D7afOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary
  • This paper studies the untargeted model poisoning in synchronous federated learning. Each poisoned client has a source model which was trained on data with corrupted labels. Upon selection, it trains a clean anchor from the current global model, aligns the source to that anchor via entropically regularised optimal transport and then submits an interpolation of the anchor and the aligned source. The attack makes use of local data and the broadcast model, without observing the updates from the honest clients or applying any form of explicit replacement amplification. The paper details transport propagation for CNNs, ResNets, vision transformers and BERT-style models. The key result is that the full transport operation leads to a large reduction in update distance whilst still preserving a large amount of the poisoning effect of unaligned interpolation.
Critical strengths
  • When a source-based comparison is carried out, the full transport operation has a significant impact. Both the linear method and OTPA align with the stored source, the clean anchor, the interpolation weight, and the schedule. The overall results provide a solid empirical basis evaluation.

  • The evaluation is broad and does provide useful limits. It involves non-IID partitions, partial participation, various architecture families, the language used, and the actual defended accuracy. The authors also mention negative results: FoolsGold gives all OTPA clients a zero weight at the final checkpoint, while FLAME manages to recover its own clean-control accuracy despite having a large number of benign rejections.

Critical weaknesses

W1. The exact improvement upon the most recent OT-based attack has not yet been clearly established. FL-MMR already incorporates OT alignment between the poisoned and benign models, and the submission refers to this in Section 2 and Table 6, so this is not an issue of missing references. The example released by the authors makes use of a fixed surrogate and applies OT following a fresh round of local training derived from the current global model. As a result, source persistence and anchor refresh on their own do not constitute a new protocol. There are indeed specific differences in OTPA's entropic transport along the feature axes and its dense interpolation. The earlier example that was examined also involves operations carried out on the server side, and therefore it does not define the same threat models or involve identical algorithms. Yet, to demonstrate the value of OTPA's particular choices, a matched comparison or a controlled comparison of the shared components with clear information and budget constraints is needed to support the claim.

W2. Figure 1 lacks better presentation and captioning, and hence paper is hard to understand at a high level.

Questions for authors
  1. During the training phase. In the case of paired full FL runs, does the uniform transport together with the control in which the correspondence has been correctly destroyed still show the attack impact of OTPA and maintain the defended performance? A typical example of a setting would help to establish whether the correspondence learned provides a meaningful advantage beyond contraction

  2. Regarding replication and defence controls, what is the number of independent runs corresponding to each interval and which random factors change? Provide, for each of the defence procedures in Table 4, the accuracy and uncertainty of the unattacked results for the decisive paired comparisons. The supplementary implementation referred to in the manuscript was not part of the material available for this evaluation.

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

This paper proposes OTPA, a federated learning model poisoning attack method based on Optimal Transport. This method performs feature alignment and parameter interpolation between the pre-trained malicious model and the normal model of the current round to reduce the parameter distance of malicious updates while maintaining the attack effect. The authors conduct experimental validation on multiple datasets, model architectures, and robust aggregation methods.

Critical strengths
  1. This paper studies the relationship between parameter distance and the actual impact of malicious models in federated learning, which has certain research value.
  2. Experiments cover multiple data sets, model architectures, and defense mechanisms, and distinguish information access assumptions for different attack methods.
Critical weaknesses
  1. Innovation is not clear enough. Existing work FL-MMR (Kumar et al., 2025) also uses Optimal Transport to align malicious models and normal models for federated learning poisoning. Although the author conducted a qualitative comparison in Appendix A, it did not fully explain the substantive innovation of OTPA in the core technical mechanism, and it also lacked direct experimental comparison with FL-MMR.

  2. The experimental verification of the core mechanism is insufficient. Appendix E.1 shows that parameter shrinkage itself can explain most of the reduction in parameter distance. However, these control experiments only evaluated the parameter distance in the model construction stage and did not compare the attack effects in full federated training, so they are not enough to prove that OT feature alignment brings significant advantages compared to simple parameter shrinkage.

Questions for authors
  1. Compared with FL-MMR, what is the main technological innovation of OTPA? Why was FL-MMR not included in the main experimental Baselines? If there are incompatibilities in experimental settings or attack hypotheses, can you explain them in detail?

  2. Can Contraction-matched Baselines be added to complete federated training to compare attack accuracy, parameter distance and defense effects to further prove the contribution of OT Alignment itself?

Reviewer_8WSrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OTPA, an untargeted model poisoning attack on federated learning. In this scenario, each malicious client keeps a poisoned model trained on noisy labels, and in every training round it participates it also trains a clean model starting from the current global model. It aligns the poisoned model to the clean one with entropic optimal transport, then interpolates between the two and submits the result. The goal is to keep most of the poisoning effect while the malicious update stays much closer to the rest of the submitted updates than plain unaligned interpolation would, without access to honest clients’ updates.

Critical strengths
  1. Extensive experimental evaluation. OTPA is compared with nine poisoning baselines and tested under six robust aggregators. It separates attacker-local baselines from those that need honest clients’ updates, which makes it clear which methods have an information advantage over OTPA. The additional experiments on different architectures, heterogeneous data, and partial participation suggest that the results are not tied to one particular setup.

  2. Weak attacker assumptions and broad applicability. OTPA only needs the broadcast model and the attacker’s own data. It does not use honest clients’ updates, private defense settings, or any change to the server protocol. This paper also describes how the transport maps propagate through convolutional, residual, ViT and BERT-style layers, so the attack is not limited to MLPs or CNNs.

Critical weaknesses
  1. Insufficient analysis of the alignment mechanism. The distance reduction come mostly from the contraction of the soft transport maps rather than from feature alignment. Appendix E.1 tests this only at the construction level, not its effect on attack strength during training.

  2. Limited effectiveness against certain defenses. Although OTPA performs well against several robust aggregators, FoolsGold and FLAME successfully identify and suppress its malicious clients. This limits the generality of OTPA's effectiveness against defenses that consider update directions or client histories.

  3. Initialization dependent advantage. Each poisoned source is trained from an independent random initialization, which introduces feature misalignment with the clean anchor. This may exaggerate the advantage of OTPA over unaligned interpolation. The paper does not evaluate the alternative of initializing the poisoned source from the global model.

Questions for authors
  1. Could the authors evaluate the correspondence free and uniform transport controls from Appendix E.1 over the full federated training process? This would help clarify how much of OTPA's advantage comes from meaningful feature alignment rather than contraction.

  2. Could the authors compare OTPA with unaligned interpolation when the poisoned source and the global model share the same initialization? This would help determine whether OT alignment is still necessary without the initial feature misalignment.

  3. Several defended accuracies in Table 4 are substantially higher than the clean FedAvg accuracy of 66.7%, such as 84.49% for norm matched noise under Krum and 81.45% for BadSampler under FLTrust. Could the authors report the clean accuracy under each defense and explain these large differences? This would help clarify how much accuracy degradation is actually caused by the attacks.

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

Related papers

Discussion (0)

Sign in to comment.