Byzantine-Resilient Federated Learning via Reverse Aggregation

Federated learning is known to achieve optimal performance, which is the highest performance achievable through central learning, only by sharing locally updated models with the central server. However, in practice, the performance of federated learning can be compromised by Byzantine failures. In this paper, we propose a new method for mitigating such backdoor attacks, namely reverse aggregation, that is robust in both data and model poisoning attacks from multiple adversarial edge devices. Unlike conventional methods that utilize the predefined number of compromised edge devices to handle the backdoor attacks, which may be unavailable in practice, proposed reverse aggregation works without any such prior information with the aid of small-sized globally shared data. Numerical results validate the advantage of the proposed method by showing improved performance compared to the conventional Byzantine-resilient federated learning methods, e.g., Krum, Trimmed mean, and Loss Function based Rejection. Further investigation compares the error rate of the proposed approach with exhaustive search, or complete enumeration method, to show that the proposed reverse aggregation has better computation efficiency with negligible performance degradation
$(\sim 0.01{\%}$
.