FedDCA: Robust Federated Learning via Fine-Grained Channel Anomaly Detection and Consensus-Constrained Aggregation

Abstract

Federated learning (FL) enables collaborative model training across decentralized data sources while preserving user privacy. However, the inherent lack of visibility into clients’ local training processes exposes FL to significant security risks, particularly backdoor attacks, in which malicious participants can stealthily embed hidden behaviors into the global model. These risks are further amplified in non-independent and identically distributed (non-IID) scenarios, where natural heterogeneity across participants can obscure the distinction between benign and malicious contributions, complicating the detection of adversarial activities and increasing the likelihood of backdoor persistence over successive training rounds. To enhance the robustness of FL systems in such complex scenarios, we propose FedDCA: Robust Federated Learning via Fine-Grained Channel Anomaly Detection and Consensus-Constrained Aggregation. FedDCA introduces a channel-wise anomaly detection module that inspects individual convolutional channels, achieving enhanced sensitivity to localized adversarial patterns while maintaining computational efficiency. For each client, channel-level deviations are quantified and aggregated into a comprehensive anomaly score, enabling the preliminary filtering of potentially malicious participants. Recognizing that anomaly detection is inherently imperfect, FedDCA further incorporates a Directional Consensus Score (DCS) mechanism, which enforces per-dimension directional consistency among clients classified as benign. This second layer of defense mitigates the influence of adversarial contributions that may evade initial filtering, thereby strengthening the robustness of the global aggregation process. Extensive experiments on multiple benchmark datasets under diverse non-IID conditions demonstrate that FedDCA effectively counters a wide range of backdoor attacks, consistently achieving lower attack success rates than state-of-the-art defenses while preserving high benign performance.

Keywords

Robust Federated Learning / Non-IID / Anomaly Detection / Directional Consensus Score

RIGHTS & PERMISSIONS

Higher Education Press 2026

Just Accepted

This article has successfully passed peer review and final editorial review, and will soon enter typesetting, proofreading and other publishing processes. The currently displayed version is the accepted final manuscript. The officially published version will be updated with format, DOI and citation information upon launch. We recommend that you pay attention to subsequent journal notifications and preferentially cite the officially published version. Thank you for your support and cooperation.