A blockchain-based federated learning framework against poisoning attacks in the internet of vehicles

Abstract

Federated Learning (FL) offers a decentralized solution for training machine learning models across distributed devices, making it well-suited for the Internet of Vehicles (IoV), where large volumes of sensitive data are generated. Despite this, FL systems are susceptible to poisoning attacks, which can compromise model integrity and performance. To address these challenges, this paper proposes SPBFL-IoV, a secure and privacy-preserving blockchain-based federated learning framework for IoV environments. The framework is specifically designed to defend against poisoning attacks, such as label-flipping and model manipulation. The proposed framework integrates blockchain technology to securely record model updates in a tamper-proof and auditable ledger, ensuring their integrity and verifiability. In addition, Homomorphic Encryption (HE) is employed to protect the confidentiality of data and model parameters during communication. Furthermore, to preserve the robustness, accuracy, and integrity of the global model in the presence of malicious participants, we employ advanced Filtering and Clipping mechanisms to identify and mitigate malicious updates. Experimental results demonstrate the effectiveness of SPBFL-IoV in terms of Overall Accuracy (All-Acc), Source-class Accuracy (Src-Acc), and Attack Success Rate (ASR), achieving an All-Acc of 98.10 % and Src-Acc of 96.00 % on the MNIST dataset, and an All-Acc of 76.15 % and Src-Acc of 60.10 % on the CIFAR-10 dataset. Furthermore, it maintains a low ASR of 0.39 % on MNIST and 9.23 % on CIFAR-10. Compared to existing methods, these results demonstrate the framework’s superior capability in countering poisoning attacks. Overall, the framework maintains high performance as measured by All-Acc and Src-Acc, and resilience against adversarial behavior, as reflected in its low ASR, making it a trustworthy solution for secure and collaborative learning within the IoV.