acceptodds
Under review as a conference paper at ICLR 2027

Risk-Guided Asynchronous Recovery for Federated Test-Time Adaptation under Distribution Shifts

Abstract

Federated Learning (FL) enables decentralized model training without sharing raw data, yet deployed models often encounter distribution shifts. To maintain model performance under such evolving shifts, federated test-time adaptation has emerged as a promising approach that adapts deployed models using unlabeled test data and periodically leverages cross-client collaboration to recover from local adaptation errors. However, as data distributions continue to shift after deployment, existing federated test-time adaptation methods suffer from: (1) errors accumulate during long-term TTA, (2) periodic collaboration incurs redundant communication, and (3) fixed layer recovery ignores degradation severity and may overwrite useful adapted parameters. To address these issues, we propose FedART, a risk-aware asynchronous recovery framework that efficiently mitigates error accumulation under continual distribution shifts. Compared to existing methods, FedART enables each client to initiate recovery asynchronously and adaptively determines recovery scope, thereby preserving beneficial local adaptations while improving efficiency. Specifically, FedART first introduces a local adaptation risk monitoring scheme to continuously estimate state risk of each client’s local TTA, and designs a risk-triggered asynchronous collaboration mechanism to retrieve suitable client adaptation states to construct recovery candidates, avoiding error accumulation and reducing redundant communication. Then, we devise a risk-guided selective layer restoration mechanism that preserves useful adapted parameters instead of uniformly recovering layers, thereby improving model stability under continual distribution shifts. Extensive experiments across multiple benchmarks demonstrate that FedART achieves competitive or superior performance compared with strong baselines, while reducing the communication volume for TTA recovery by approximately 98% compared with periodic recovery.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Test-time Adaptation, continual distribution shifts, continual learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_6A1aOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

To address challenges in federated test-time adaptation under continuous distribution shifts,specifically, when a client should recover, which clients to consult for assistance, and how many parameters to restore, the authors propose a recovery framework called FedART. FedART monitors risk using uncertainty signals calibrated against the source domain; upon triggering an alert, it retrieves client states that are similar, recent, and estimated to be reliable. It then selects recovery layers based on prediction concentration and employs a peer-prediction consensus mechanism to compare candidates across normal adaptation, local recovery, collaborative recovery, and rollback strategies. In my view, this work effectively resolves the robustness issues inherent in TTA-FedDG through the clever use of specific technical strategies.

Critical strengths
  1. The author addresses a practical and complex problem, focusing on mechanisms of degradation and recovery during long-term test adaptation, while considering multidimensional factors such as client-side asynchronous drift, error accumulation, and communication overhead. The study specifically investigates "when to collaborate" and "how much to recover."

  2. The authors provide a comprehensive treatment involving both experiments and theoretical analysis, accounting for various factors such as client heterogeneity, availability, population size, and state staleness; this demonstrates the system's robustness to a certain extent. Furthermore, the most notable feature of this work is its extremely low communication cost.

Critical weaknesses

1.Although the authors report average accuracy and worst-client accuracy, it remains difficult to characterize short-term performance collapses and recovery capabilities during the adaptation process. To demonstrate the method's robustness, the authors should provide additional metrics such as the miss rate, worst-case time-window accuracy, frequency of performance collapses, recovery time following degradation, and the proportion of recovery operations that actually improve the target accuracy.

2.The authors proposed this method to address the poor robustness associated with the ResNet model architecture; they should also provide results demonstrating the method's performance on pre-trained ViT models.

Questions for authors
  1. The author's figures need to be redrawn; some of the text and graphics appear flattened, which affects readability.

2.Refer to the issues mentioned in the weakness. I will lower the score if the author does not address my questions.

Reviewer_238jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FedART for federated test-time adaptation under distribution shifts. Clients monitor prediction uncertainty and retrieve suitable peer states asynchronously when a risk alarm is triggered. The method adjusts restoration depth based on prediction concentration and selects recovery candidates using peer agreement. Experiments on CIFAR-10-C, CIFAR-100-C, PACS, and OfficeHome report competitive accuracy with substantially reduced recovery communication.

Critical strengths
  1. FedART combines risk monitoring, asynchronous peer retrieval, and selective parameter restoration in a unified framework. This provides a concrete approach to reducing unnecessary recovery communication while retaining useful local adaptation.

  2. Experiments on four benchmarks show competitive or improved adaptation accuracy with substantially lower recovery communication. The component ablations also support the contribution of the three main mechanisms.

Critical weaknesses
  1. The restoration policy in Section 3.5 is somewhat heuristic. The paper does not sufficiently justify using prediction concentration to determine restoration depth or restoring the deepest BN-affine layers first. Concentration can also change with class composition. The structural analysis in Appendix A.3.2 does not establish the accuracy benefits of these choices.

  2. The reliability of the risk monitor in Section 3.3 needs further validation. It relies on the evolving model's confidence and a source-calibrated proxy assumption, leaving unclear how effectively it detects confidently incorrect predictions. Missed alarms and detection delays are particularly relevant to the claimed recovery benefits.

  3. The evidence linking recovery to reduced error accumulation is indirect. Figure 4 shows temporal accuracy but does not identify alarms, accepted recoveries, or their subsequent effects. Relating these events to target error would help establish whether the mechanism interrupts degradation rather than merely producing better overall accuracy.

  4. The benefit of risk-guided timing is not sufficiently isolated. Comparing trigger schedules with identical peer selection, restoration, and commit rules would distinguish better timing from simply performing fewer recoveries. Communication accounting should also clarify whether registry maintenance and peer-state publication are included.

  5. The local TTA procedure is underspecified in Section 3.1 and Appendix A.2. The abstract operator T is not instantiated with a specific adaptation loss, optimizer, learning rate, or number of updates per batch. These details are necessary for reproduction and for assessing whether the recovery methods use comparable local adaptation.

Questions for authors

See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.