Reproduce “Deep Federated Learning for Autonomous Driving” on OpenResearch

About this paper

Anh Nguyen, Tuong Do, Minh Tran, Binh X. Nguyen, Chien Duong, Tu Phan, Erman Tjiputra, Quang D. Tran

Researchers from the University of Liverpool and AIOZ developed a peer-to-peer Deep Federated Learning (DFL) framework combined with a custom neural network, FADNet, to enable privacy-preserving training of end-to-end autonomous driving policies. The approach achieved competitive or superior accuracy compared to centralized and server-based federated learning methods across multiple datasets, with a real-world deployment demonstrating practical viability.

Key insights

• Peer-to-peer DFL can achieve competitive or superior performance for complex tasks like autonomous driving, outperforming server-based FL by mitigating communication bottlenecks and single points of failure.
• Custom neural network architectures like FADNet are crucial for stable convergence and high accuracy in federated learning, particularly when dealing with non-IID and imbalanced data distributions.
• The approach offers enhanced privacy by keeping raw data local and demonstrates scalability and robustness across different network topologies and increasing numbers of participating clients.

Summary by alphaXiv.