AsyncDefender: Dynamic trust adaptation and collaborative defense for Byzantine-robust asynchronous federated learning

Abstract

Traditional synchronous federated learning approaches face efficiency challenges in resource-heterogeneous networks and high-latency environments. The appearance of asynchronous federated learning has solved some of these limitations. However, Byzantine fault-tolerant approaches that work well in synchronous environments face unique challenges in asynchronous federated learning, such as synchronous robust rules that cannot be applied to asynchronous scenarios and difficulty in coping with dynamic changes. To address these issues, we introduce AsyncDefender, one of the first solutions designed specifically for Byzantine fault tolerance in asynchronous federated learning.AsyncDefender supports fully asynchronous updates from edge clients, tolerates arbitrary latency, and has no limitation on the number of Byzantine clients. The core of our method is the dynamic allocation of aggregation degrees based on the similarity between client gradients and global model gradients. In addition, we employ collaborative filtering between reviewers and non-reviewers and bidirectional credibility assessments to identify and eliminate malicious updates accurately. Extensive qualitative and quantitative experiments demonstrate that AsyncDefender is not only more robust to a large number of malicious Byzantine clients but also converges faster and performs more stably than existing approaches.