Secure Aggregation Techniques in Federated Learning for Vehicle Data Analytics

The rapid evolution of Intelligent Transportation Systems (ITS) and Autonomous Vehicles (AVs) has generated a massive influx of vehicular data. While this data is pivotal for enhancing traffic safety and predictive maintenance, privacy concerns regarding location history and driving patterns remain a significant barrier. Federated Learning (FL) offers a decentralized alternative to traditional machine learning by training models locally on vehicles; however, FL is still susceptible to poisoning attacks and inference-based privacy leaks. This study investigates the efficacy of secure aggregation techniques—specifically Homomorphic Encryption (HE) and Multi-Party Computation (MPC)—in maintaining model accuracy while ensuring robust privacy. Using a quantitative experimental design and the Bosch Vehicle Motion Dataset, we demonstrate that while secure aggregation introduces a latency overhead of 12-18%, it successfully mitigates reconstruction attacks without compromising convergence rates. Our findings suggest that a hybrid approach is optimal for real-time vehicular analytics.