YOLO Meets FedAVG: A Privacy-Preserving Approach to Autonomous Vehicles Object Detection

Autonomous vehicle (AV) systems rely heavily on accurate object detection to ensure safe and efficient operation, with models like YOLO excelling in real-time performance and precision. However, in the context of AV applications, traditional centralized object detection methods often struggle with data privacy concerns and the need for generalization across diverse, dynamic environments. Existing research predominantly focuses on centralized approaches, which lack sufficient privacy measures and face limitations in scalability when applied to decentralized, real-world settings. To address these challenges, we propose a privacy-preserving framework that integrates YOLOv8 with the Federated Averaging (FedAVG) algorithm, allowing decentralized training across multiple AV units while retaining data privacy. This approach trains the model locally over 10 epochs per client, followed by aggregation over 10 communication rounds, enhancing detection accuracy while meeting privacy requirements critical to AV deployment. Experimental evaluations demonstrate the effectiveness of our YOLOv8-FedAVG framework, achieving a mean Average Precision (mAP) of 86%, precision of 84%, and recall of 83%, showcasing its potential to advance AV object detection in a scalable, privacy-focused manner.