Anonymous and Efficient Authentication Scheme for Privacy-Preserving Federated Cross Learning

Federated learning (FL) reduces privacy risks by allowing participants to train models locally without transmitting raw data to a central server. However, existing research suggests that adversaries are able to obtain data information by analyzing successive model updates, and can also violate privacy by inferring a link between the data and the client. The most advanced solutions to the problem of privacy leakage are mainly differential privacy (DP) and cryptography techniques. Whereas, the implementation of DP tends to degrade data utility, while cryptographic techniques often incur prohibitive computational costs and communication overheads. To solve these challenges, we propose a ring signature anonymous authentication scheme based on elliptic curve cryptography, which ensures high computing efficiency and data utility and privacy security. We provide formal security proofs and analysis. Moreover, performance evaluations illustrate the superior computational efficiency of our proposed protocol. Tailored encryption for parameter cross-training and tracking mechanisms for identifying malicious actors are proposed to align with federated cross learning (FedCross). Comprehensive empirical evaluations demonstrate that the anonymous authentication scheme does not compromise model performance and provides robust privacy safe-guards.