Efficient privacy aggregation method based on zero-knowledge proofs in federated learning
Zheyu Li, Jiangfeng Xu, Bo Wang · 2024
Federated learning (FL) enables distributed clients to privately train machine learning models using their own local data, thus avoiding the security risk of directly exchanging private data between clients and servers. However, there is a risk in federated learning: a malicious party may launch a reverse attack based on the parameters uploaded by the client to analyze the attributes of the client’s local private data. Some studies use the Shamir secret sharing method and zero-knowledge proofs (ZKP) to ensure the privacy and integrity of the inputs in FL. However, the Shamir secret sharing scheme often requires an optimistic guarantee on the number of malicious clients, which cannot obtain a sufficient number of secret slices through collusion, and the general ZKP scheme requires the clients to compute the proof belongs to others during the proof and verification stages, and its efficiency needs to be improved. In this paper, we propose VSSPFL method to achieve efficient and secure data collaboration, which improves the Shamir secret sharing scheme to prevent reverse attacks by an unknown number of malicious clients and reduces the cost of ZKP by using a probabilistic integrity check method.