PureFL: Noise-Resilient Homomorphic Encryption for Blockchain-Based IoV Federated Learning
Paper Title: PureFL: Noise-Resilient Homomorphic Encryption for Blockchain-Based IoV Federated Learning
Journal Name: Artificial Intelligence Review
Abstract: Federated learning (FL) in the Internet of Vehicles (IoV) networks is challenged by privacy vulnerabilities, model poisoning attacks, and the limitations of centralized aggregation, which threaten security and scalability. While homomorphic encryption (HE) provides an additional layer of privacy, existing schemes often suffer from cumulative noise growth, resulting in degraded model accuracy and increased inference latency, critical factors for real-time vehicular safety applications. This paper presents PureFL, a novel FL framework that combines adaptive Cheon-Kim-Kim-Song (CKKS) homomorphic encryption with a custom permissioned blockchain (PureChain), utilizing a proof of authority and association (PoA^2) consensus mechanism. PureFL incorporates dynamic precision scaling to manage noise accumulation, ensuring computational efficiency without sacrificing privacy or accuracy. The framework also implements a trust-based aggregation mechanism based on cosine-similarity trust scoring, which dynamically weights client contributions according to historical reliability and anomaly detection, thereby enhancing resilience against model poisoning. Experimentation results on the CICIOV2024 and 5G-NIDD datasets demonstrate that PureFL achieves superior anomaly detection accuracy of 99.9%, reduced CKKS HE overhead with precision scaling, and the proposed PoA^2 consensus demonstrates superiority to other consensus mechanisms in terms of throughput and latency, making it suitable for real-time IoV operations.