Byzantine-Robust Decentralized Federated Learning via Local Performance Checking

Authors

Kaichuang Zhang, The University of Texas Rio Grande Valley
Alina Basharat, The University of Texas Rio Grande Valley
Ping Xu, The University of Texas Rio Grande Valley

Document Type

Conference Proceeding

Publication Date

2025

Abstract

Decentralized federated learning (DFL) offers enhanced resilience to client failures compared to its centralized counterpart, due to its ability to aggregate learning models from distributed clients without the need for centralized server coordination. However, the presence of malicious Byzantine clients can degrade the performance of DFL algorithms, or even cause the entire learning process to fail. To combat Byzantine clients in DFL, this paper proposes a straightforward yet efficient algorithm to detect malicious updates using a small dataset sampled from clients’ local dataset. The proposed Local Performance Checking (LPC) algorithm allows each client to calculate the performance metrics (e.g., classification accuracy) of received updates from their neighbors using a pre-sampled performance test dataset. Updates whose performance metrics deviate significantly from those of the local model’s are identified and filtered out. The effectiveness of the proposed algorithm is demonstrated through extensive simulations across a wide range of practical scenarios.

Comments

© 2025 The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.

Recommended Citation

Zhang, Kaichuang, Alina Basharat, and Ping Xu. "Byzantine-robust decentralized federated learning via local performance checking." In International Conference on Neural Information Processing, pp. 171-185. Singapore: Springer Nature Singapore, 2024. https://doi.org/10.1007/978-981-96-6954-7_12

First Page

171

Last Page

185

Publication Title

International Conference on Neural Information Processing

DOI

10.1007/978-981-96-6954-7_12