Byzantine-Robust Hierarchical Aggregation for Cross-Device Federated Learning in Consumer IoT

Abstract

Nowadays, Federated Learning (FL) has emerged as a prominent technique of model training in Consumer Internet of Things (CIoT) without sharing sensitive local data. Targeting privacy leakage of cross-device FL in CIoT, various privacy-preserving FL schemes have been proposed. Regrettably, existing schemes still face three significant challenges: 1) Current privacy-preserving strategies struggle to fully defend against Byzantine attacks in FL without compromising data privacy; 2) Most privacy-preserving techniques (e.g., secret sharing) in FL result in substantial computation and communication overhead; 3) The non-colluding dual-server setting limits the applicability of FL. To overcome these challenges, we propose a Byzantine-robust hierarchical federated learning scheme, named BHFL. This scheme not only effectively defends against Byzantine attacks while safeguarding user privacy but also avoids the need for a dual-server architecture. Simultaneously, the hierarchical aggregation structure can effectively train non-IID cross-device data while maintaining high communication efficiency. We evaluate BHFL on several benchmark datasets, and the experimental results demonstrate that BHFL achieves high accuracy and Byzantine robustness compared to the popular FedAvg scheme. Therefore, BHFL is well-suited for CIoT scenarios.