FEDERATED LEARNING IN AUTONOMOUS VEHICLES: BALANCING ACCURACY, PRIVACY, AND LATENCY

The integration of federated learning (FL) into autonomous vehicles (AVs) offers a promising path toward privacy-preserving, decentralized, and intelligent mobility systems. However, real-world deployment of FL in AV environments presents critical trade-offs among model accuracy, data privacy, and communication latency. This study investigates these trade-offs by simulating three distinct FL configurations: baseline FL, differential privacy-enhanced FL, and communication-optimized FL. Using public datasets that emulate AV sensor data, each configuration was evaluated on the basis of model performance, privacy risk, and system responsiveness. Results showed that while baseline FL yielded the highest accuracy (92.3%), it exposed sensitive data. Differential privacy-enhanced FL ensured the lowest privacy risk (score: 0.3) but suffered from increased latency and reduced accuracy. Communication-optimized FL achieved the best overall balance, with competitive accuracy (89.9%), acceptable privacy, and the lowest latency (112 ms). A composite evaluation revealed that no single configuration excels in all areas, highlighting the need for adaptive, context-aware FL strategies. The findings underscore the importance of system-level optimization in deploying FL for autonomous driving, where safety, responsiveness, and data protection must coexist.