Efficient and Privacy-Preserving Authentication for Federated Learning in Industrial Internet of Things Data Sharing Application

Federated learning (FL) has emerged as a promising approach for collaborative machine learning while respecting data privacy. However, directly aggregating local models during FL training can expose sensitive information, as adversaries may analyze model parameters to infer the underlying raw data. Therefore, a robust authentication mechanism is required to verify the integrity of local models and confirm their source. However, many existing authentication schemes are computationally intensive, which can hinder performance. In this paper, we present a lightweight authentication scheme for FL called EPPAFL. This innovative approach is compatible with all machine learning algorithms, maintains data utility, and ensures strong privacy protection. Additionally, it can be integrated with detection mechanisms to defend against data poisoning and free-rider attacks. Specifically, we first design a certificate-less signature scheme based on elliptic curve cryptography (ECC) and then propose an anonymous authentication protocol based on the Chinese remainder theorem (CRT), which supports batch verification, dynamic updates, and revocations of group keys. Moreover, participants interact anonymously for secure communication, and a trusted authority (TA) can trace and recover their identities in case of malicious activities. We provide formal security proofs and analyses to demonstrate the achieved security properties. Performance evaluations demonstrate that our proposed protocol significantly reduces the costs associated with signature generation and verification, outperforming existing schemes in both communication and computational efficiency.