FedCom: Byzantine-Robust Federated Learning Using Data Commitment

Bo Zhao, Tao Wang, Liming Fang

Abstract: Federated learning is a promising distributed edge learning methodology that allows multiple clients to collaboratively train statistical models without disclosing private training data. However, there may exist Byzantine clients launching data/model poisoning attacks to compromise global model's performance or convergence. Most of the existing Byzantine-robust FL schemes are either ineffective against several advanced poisoning attacks, and their robustness suffers from further degradation when local datasets are highly non-independently and identically distributed (non-IID). To address these issues, we propose FedCom, which could achieve data/model poisoning tolerant FL under practical non-IID data partitions, even the attackers do not honestly follow FedCom's protocol. The cardinal design of FedCom is privacy-respectfully asking clients to generate the commitments, which commit and verify the honesty of their local data distributions or local model updates. An extensive performance evaluation demonstrates FedCom's superior performance compared to the state-of-the-art Byzantine-robust schemes under various rigorous settings, even the attackers do not honestly follow the protocol of FedCom and fabricate the commitments.