Federated Learning Based Privacy Preservation in Autonomous Driving
Abstract
Autonomous Driving (AD) is the technology that enables a vehicle to navigate itself from a source to a specified destination with minimal human involvement, safely and intelligently. To achieve it in real-world scenarios, various integrated AD systems have been developed with a combination of sensors like camera, LiDAR, GPS, artificial intelligence (AI), and complex software working together to perceive the vehicle's environment and drive behaviors, make real-time decisions, and physically control the car's movement (steering, accelerating, and braking) autonomously. IT refers to data-driven, sensor-enabled frameworks that allow vehicles to perceive their surroundings, plan actions, and execute control. However, these data-driven, sensor-enabled AD systems heavily rely on the continuous collection and processing of sensitive information, posing substantial privacy risks at both the individual and population levels. Addressing these risks requires a new learning framework that minimizes data exposure while preserving model accuracy and scalability. This study investigates Federated Learning (FL) combined with Differential Privacy (DP) as a privacy-preserving alternative to centralized machine learning for AD. Three datasets are introduced and leveraged in our methodology to support driving environment modeling, object classification, perception, and end-to-end AD evaluation. Specifically, we utilize CIFAR-100 (Dataset 1) and its feature-engineered subset, CIFAR-36 (Dataset 2), for the classification tasks, while Dataset 3 is paired with the CARLA simulator and PlanT architecture for closed-loop AD evaluation. Results show that FL consistently outperforms isolated local learning and approaches the accuracy of centralized baselines, all while ensuring raw data remains locally. Although modest DP noise preserves model efficacy, excessive noise substantially degrades accuracy and can impede effective learning. The findings in this study demonstrate that FL, combined with carefully calibrated DP mechanisms, provides a practical framework for privacy-preserving model training in AD, reducing the need to transfer sensitive data to Central Server.
