Federated Learning for Autonomous Vehicles : Privacy-Preserving Edge AI

Abstract

Federated Learning (FL) offers an appropriate approach to utilizing the huge amount of data created by Autonomous Vehicles (AVs). Besides this ability, it focuses on preserving data privacy while working on it. This thesis conducts a thorough, review-based analysis of the use of federated learning in autonomous vehicles as a privacy-preserving method of edge AI. The discussion is initiated by identifying the primary technical and ethical challenges. These issues include data heterogeneity across different driving situations, limitations in vehicle resources, and the necessity for fairness and transparency in model updates. This work examines various privacy-enhancing techniques—Differential Privacy, Secure Multi-Party Computation, Homomorphic Encryption, and Secure Aggregation—to evaluate their trade-offs in protecting sensitive information while preserving model utility. The survey then analyses communication solutions specifically designed for automotive networks, including model compression techniques (Quantization, Pruning) and asynchronous updating methodologies, as well as Hierarchical and Decentralized Federated Learning architectures. Moreover, the survey tries to emphasize the effectiveness of these techniques in mitigating bandwidth and latency constraints. One of the significant purposes of this survey is to emphasize the issue of non-IID data. To address this issue, the survey examines techniques such as FedProx, SCAFFOLD, and Clustered FL that are beneficial to resolve client drift and enhance convergence when local training datasets exhibit distributional discrepancies. Another topic that has been addressed in this thesis is personalization strategies, including Transfer Learning and client-specific adaptation. These techniques are helpful to balance global knowledge with local complexity. The survey explores existing security concerns, which include adversarial and Byzantine attacks, and tries to mention some mitigating strategies through robust aggregation and blockchain-supported federated learning frameworks. The results of these investigations lead to specific recommendations: the necessity for communication-efficient client selection, adaptive privacy controls tailored to vehicular situations, realistic simulation benchmarks (e.g., FLEXE), and standardized industry protocols. This thesis clarifies the transformational potential of FL for AVs by integrating several elements and identifies research directions, encompassing algorithms, systems, and ethics, that influence the next generation of safe, efficient, and privacy-conscious autonomous transportation.