Secure Distributed Machine Learning Client Selection Algorithm Based on Privacy Leakage Weight
Yuanhao Xu, Shihao Zhang, Yifan Ding, Zihao Wang · 2024
As a secure distributed machine learning framework, federated learning is often used for secure data transmission. However, existing research has found that federated learning also faces the risk of privacy leakage, such as gradient leakage attacks. Existing federated learning algorithms mainly select participating clients through random selection. These federated learning algorithms assume that all clients have the same risk of being attacked. However, existing research has found that due to differences in the local model data distribution of different clients, the resistance of different clients to attacks varies. Traditional random client selection methods cannot better quantify the resistance of clients to attacks, thus failing to guarantee the privacy security of the federated learning process. In this paper, we propose a secure federated learning client selection algorithm based on privacy leakage risk assessment. We calculate the privacy leakage weight of each client by measuring the privacy leakage risk of each client, using it as the criterion for selecting clients. Furthermore, we dynamically select clients in an attempt to select clients with smaller privacy leakage weights to participate in federated learning training, thereby ensuring the security of our data privacy.