Federated Learning with Data-Free Distillation for Heterogeneity-Aware Autonomous Driving

Junyao Liang, Juan Li, Ji Zhang, Tianzi Zang

Abstract: Autonomous driving is a hot topic within both academic and industrial domains. Critical to its advancement is the high performance of used machine learning models such as image recognition, lane detection, and danger prediction models. Sharing daily driving data is essential for the continuous enhancement of these predictive models. However, a significant challenge arises in balancing data sharing with the need to protect sensitive driving information. Federated learning (FL) is a popular mechanism for learning models from different vehicles without leaking private local data. However, FL often lacks robustness, leading to suboptimal models when local data comes from highly heterogeneous distributions. In this paper, we propose a novel FL framework with data-free knowledge distillation to address the above problems. Each vehicle extracts local hyper-knowledge including intermediate layer representations and soft predictions during local training. The hyper-knowledge is transferred to the FL server and aggregated to guide the next round of training. By this way, we can acquire better local models that perform well w.r.t. corresponding vehicle’s distribution and a better global model that can be used by future new vehicles without history driving data. Compared with existing knowledge distillation methods, we do not need public data to distill knowledge, which avoids privacy leakage when a real public dataset is shared and high data generation costs when a virtual dataset is shared. The experiments based on traffic sign recognition verify our advancement compared to baselines.