Privacy-Preserving and Robust Federated Learning Based on Secret Sharing

Federated learning (FL) is a machine learning method that enables model training without centralizing data for integration. However, FL is vulnerable to poisoning attacks, in which an attacker manipulates the malicious clients to corrupt the global model via poisoning their local training data or model updates, resulting in compromised model accuracy and degraded performance. In addition, in FL, although the original data can be trained without leaving the local devices, some attackers can obtain the private information of training participants through model parameters, causing privacy leaks. In order to solve the above problems, we propose a privacy-preserving federated learning robust aggregation scheme based on secret sharing. This scheme is implemented based on secret sharing technology, protecting clients' data privacy while achieving Byzantine-robust. Moreover, our scheme considers the two situations of honest majority and malicious majority of clients; that is, the model can effectively resist poisoning attacks when the proportion of malicious clients is less than 50% or more than 50%. Extensive experiments show that our scheme is secure against various common poisoning attacks and is more robust than some existing aggregation rules, even when malicious actors account for the majority.