A decentralised asynchronous federated learning framework for autonomous driving

Traditional autonomous driving usually requires a large number of vehicles to upload data to a central server for training. However, collecting data from vehicles may violate personal privacy as road environmental information contains geographic location. Federated learning can achieve multi-vehicle collaborative sensing of the road environment while protecting data privacy. However, the existing centralised federated learning architecture faces some challenges, such as credibility, fairness, and real-time. To address the above issues, we propose a decentralised asynchronous federated learning framework based on blockchain. Firstly, using blockchain to replace the central server of traditional federated learning architecture avoids the untrustworthy issues caused by the central architecture. Secondly, the blockchain module includes scoring contract units and incentive contract units to prevent malicious vehicle attacks and designs fair incentive mechanisms to ensure the ecological health and sustainable development of federated learning. Thirdly, using the asynchronous federated learning algorithm, blockchain can immediately aggregate model updates from vehicles, greatly improving the overall training flexibility and real-time performance. Experimental results demonstrate the effectiveness of the proposed framework. Copyright © 2023 Inderscience Enterprises Ltd.