acceptodds
Withdrawn from ICLR 2027

FedHRA: Client-Side Purification and Hierarchical Robust Aggregation for Federated Backdoor Defense in Time-Series Classification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated learning enables industrial sites to train time-series classification models without sharing raw process data, but it exposes training to backdoor poisoning from compromised clients. Existing robust aggregation defenses, designed for generic cross-device settings, become unreliable under the strong non-IID distributions of industrial cross-silo FL, where benign updates naturally deviate from the global average. We propose Federated Backdoor Defense with Client-Side Purification and Hierarchical Robust Aggregation (FedHRA). On the client side, a conditional variational autoencoder learns the class-conditional normal manifold of local time-series signals; prediction flipping, confidence drop, and reconstruction error are fused into a suspiciousness score that purifies risky samples and provides client-level risk evidence. On the server side, FedHRA separates the backbone from the classification head and, using client-reported risk signals and temporal memory, applies stronger sanctions to the decision-sensitive head while softly downweighting the backbone. Experiments on the Tennessee Eastman Process and Spoken Arabic Digits datasets show that FedHRA suppresses terminal attack success rates to near zero under single-shot, continuous, periodic, and late-stage attacks while maintaining usable main-task accuracy.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Time-Series Classification, Backdoor Defense, Client-Side Data Purification, Hierarchical Robust Aggregation, Non-IID Data

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

This paper proposes FedHRA, combining client-side CVAE purification with server-side hierarchical robust aggregation to defend against backdoor attacks in non-IID federated time-series classification. Experiments demonstrate low attack success rates under different attack schedules. Although the paper addresses a relevant problem and reports promising results, concerns regarding the threat model, methodological justification, and comparisons with recent work substantially weaken its contributions.

Critical strengths
  1. The paper addresses an important security challenge in federated learning, where statistical heterogeneity can make benign and malicious client updates difficult to distinguish.
  2. The proposed framework combines sample-level purification with server-side robust aggregation, providing a relatively comprehensive defense pipeline.
  3. The experimental results demonstrate promising backdoor suppression under several attack schedules, particularly continuous and late-stage attacks.
Critical weaknesses
  1. Unreliable client-reported risk information. The introduction states that compromised clients may manipulate local training and model updates. However, FedHRA relies on clients to perform purification and report risk statistics, which malicious clients could easily bypass or falsify. This raises concerns about the trust assumptions underlying the defense.
  2. Insufficient justification for CVAE-based purification. The CVAE is pretrained on potentially poisoned local data and may therefore learn and preserve backdoor triggers. The paper lacks sufficient evidence that the learned manifold is clean or that reconstruction reliably removes backdoor patterns.
  3. Insufficient justification for hierarchical aggregation. The authors impose stronger penalties on classification heads, although backdoor functionality may also reside in the backbone. Proposition 2 does not establish that heads are more vulnerable to backdoors or that stronger head suppression is preferable.
  4. Limited motivation and outdated baselines. The three motivating challenges are common to general FL rather than specific to time-series classification. Moreover, recent non-IID federated backdoor defenses are insufficiently discussed, and the experimental baselines are limited to methods from 2023 or earlier, weakening the novelty and empirical comparisons.
Questions for authors
  1. How does FedHRA ensure the reliability of client-reported risk statistics when malicious clients can manipulate local training? The authors should clarify the trust assumptions and consider attacks that falsify these statistics.
  2. Could the authors provide evidence that the CVAE learns a clean class-conditional manifold rather than preserving backdoor patterns when pretrained on potentially poisoned local data?
  3. Could the authors further justify why classification heads require stronger suppression than backbones, both theoretically and empirically? Comparisons with uniform or reversed hierarchical weighting would help validate this design.
  4. Could the authors clarify the time-series-specific motivation and include comparisons with recent federated backdoor defenses designed for non-IID settings?
Reviewer_EYSrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. FedHRA combines client-side data purification and server-side hierarchical aggregation to defend federated time-series classifiers against backdoor attacks under non-IID data.
  2. Clients use a class-conditioned autoencoder to identify suspicious samples, while the server combines reported risk statistics, classification-head drift, and temporal memory to assign aggregation weights.
  3. Experiments on TEP and Spoken Arabic Digits report low terminal attack success rates across four attack schedules, with a trade-off in clean accuracy.
Critical strengths
  1. Relevant problem and complementary design. Combining sample purification with differentiated backbone and head aggregation addresses both poisoned training data and residual malicious updates.
  2. Broad baseline comparison. The evaluation includes numerous robust aggregation defenses, four attack schedules, and sensitivity analyses over poisoning and malicious-client ratios.
  3. Useful component analysis. Ablations and aggregation-weight visualizations help explain how client-side purification and hierarchical aggregation contribute to security and utility.
Critical weaknesses
  1. The trust model is insufficiently specified. The defense relies on potentially compromised clients to execute purification and honestly report risk statistics. An attacker could bypass purification or falsify these reports; the paper should clarify the trusted components and evaluate such adaptive behavior.
  2. The evaluation leaves important robustness questions unanswered. All attacks begin after CVAE pretraining, effectively providing a clean initialization period. Robustness to poisoning during pretraining, different trigger constructions, and defense-aware attacks is unclear. Reporting only terminal attack success also misses transient vulnerabilities, and the tables provide no variability estimates.
  3. The theoretical support is limited relative to the claims. Malicious-weight suppression assumes that malicious clients already receive higher risk scores, leaving reliable detection unproven. The purification guarantee concerns deletion, whereas experiments use replacement; moreover, TEP single-shot clean accuracy falls from 90.87% without defense to 82.05%, despite identical terminal attack success of 0.42%.
Questions for authors

Please refer to the weakness

Reviewer_XWP2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies backdoor defense in industrial FL to solve time-series classification problems. The authors argue that robust aggregation fails under strong non-IID data, because benign updates also deviate from the global average. So they move part of the defense to the client side. FedHRA combines client-side sample purification with server-side hierarchical aggregation. It uses heavy sanctions on the classification head, light touch on the backbone, and a client specific history, and it beats 12 baselines on TEP and SAD under four attack schedules.

Critical strengths
  1. The motivation is clear. Under non-IID situations, benign clients do look like outliers to server-side filters, and Fig. 1 makes this point well. This matters because it explains why the defense has to go beyond the server.
  2. Purifying samples before they become local updates is a reasonable and cheap idea.
Critical weaknesses
  1. The accuracy cost is large and unexplained. On TEP under single-shot attacks, FedHRA gets MTA 82.05 while No defense gets 90.87. There is no run with no malicious client, so this cost is never separated from the attack.
  2. Prediction probability drop, MSE reconstruction error, and a binary flip indicator are weighted together with λ₂=10 without normalization. q, ρ_susp, ρ_blk, κ, W are listed in Algorithm 1 but never reported. Besides, there is no sensitivity analysis.
  3. The threat model is too favorable. CVAE is pretrained at round 5, all attacks start at round 20+. If poisoning starts from round 1, the CVAE learns the trigger as normal and purification breaks. This case is not tested.
  4. The theory is not connected to the experiments. Proposition 1 is Bayes' rule and needs TPR and FPR, but neither is reported. Theorem 1 assumes a score gap Δ_h that is not measured.
  5. The ablation does not match the story. Removing hierarchical aggregation drops MTA to 73.69 on TEP and 49.34 on SAD, while removing client defense keeps MTA at 89.08 but raises TASR to 97.83. So the client side gives both the security and the accuracy loss. The paper also claims to protect long-tail clients, but reconstruction error and head drift are naturally high for clients with rare conditions.
Questions for authors
  1. Can you report MTA and TASR with no malicious client at all? How much of the 9-point drop on TEP comes from purifying clean samples?
  2. Please give the values of q, ρ_susp, ρ_blk, κ, and W, and add sensitivity analysis.
  3. What happens if the malicious client reports s_k = 0, or if poisoning starts at round 1 so the CV AE also learns the trigger? Does FedHRA still hold?
  4. Please report the purification TPR and FPR, since Proposition 1 depends on them, and measure the gap Δ_h assumed in Theorem 1.
  5. Are the reported numbers from a single run? Please add multiple seeds and confidence intervals.

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

Related papers

Discussion (0)

Sign in to comment.