Blockchain-based federated learning algorithm and incentive mechanism enhancing data privacy and training accuracy

Abstract

With rapid advancements in computing, communication, and storage technologies, innovative machine-learning training methods have emerged. One prominent approach is federated learning, which enables collaborative model training without sharing original data sets and ensures data security and privacy. However, federated learning faces challenges, including low-edge node participation, untrustworthy nodes, and untraceable training data. This research combines federated learning theory with blockchain technology, proposing a hybrid blockchain-based federated learning algorithm and incentive mechanism. The primary contributions are the following: Firstly, existing federated learning algorithms hide training data, making them vulnerable to backdoor attacks. To address this, the proposed algorithm utilizes a federated blockchain for authentication and management, preventing impersonation and false data injection, thereby improving training accuracy. Secondly, participating node identities are authenticated using a consortium blockchain, ensuring data integrity. Training parameters are stored on a public blockchain, enabling training data traceability. Blockchain integration enhances decentralization, security, and privacy. Simulation experiments demonstrate the proposed scheme’s superiority in robustness and accuracy compared to traditional methods. This research provides valuable insights for advancing federated learning, addressing challenges, and enhancing its applicability in real-world scenarios.