CAPFL: Context-Aware Adaptive Privacy-Based Federated Learning in Connected Autonomous Vehicles using LDP

With the rapid development of wireless technologies such as 6G and more sophisticated vehicular sensor networks used in Connected Autonomous Vehicles (CAVs), Intelligent Transportation Systems (ITS) could be collaboratively trained to effectively utilize these latest technologies. Nevertheless, this new paradigm of collaborative learning over the edge often includes sharing raw vehicular data, resulting in an increased threat to individual users’ privacy. The application of Federated Learning (FL) and Local Differential Privacy (LDP) enables CAVs to collaboratively train models without sharing raw data. However, it has been identified that an inflexible privacy budget for heterogeneous driving scenarios cannot effectively counteract several privacy attacks. In this paper, we propose a new approach for privacy budget adaptation based on federated learning, namely Context-aware Adaptive Privacy-Based Federated Learning (CAPFL). The proposed approach is based on dynamically adjusting the privacy budget based on the characteristics of the participating agents. In this paper, we also introduce privacy budget binning with mild dithering for counteracting side-channel attacks. The experiments were conducted using the Udacity dataset with a matched cumulative privacy budget for each iteration. The results showed that CAPFL improves the privacy-utility trade-off by focusing more on privacy where it is more needed, while still achieving or even enhancing accuracy and convergence with negligible system overhead.