Privacy-preserving federated learning with adaptive model aggregation for efficient vehicle-to-vehicle (V2V) communication in intelligent transportation systems

Intelligent Transportation Systems (ITS) demand robust privacy-preserving frameworks that maintain efficiency and adaptability in dynamic Vehicle-to-Vehicle (V2V) networks. Conventional federated learning (FL) approaches falter under non-IID data distributions, adversarial threats, and rapidly changing traffic conditions. This paper introduces FLAA-V2V, a novel FL framework that addresses these challenges through three key innovations: (1) A reinforcement learning-based adaptive aggregation engine dynamically weights vehicle contributions using context-aware metrics (data quality, network stability), reducing communication overhead by 23% versus FedAvg; (2) A hierarchical privacy mechanism combining Local Differential Privacy (LDP) and Lightweight Homomorphic Encryption (LHE) secures V2V exchanges while achieving 92.3% collision-avoidance F1-score under attacks; and (3) A meta-learning drift detector with Kolmogorov-Smirnov validation and gradient compensation reduces accuracy degradation by 18.7% in non-stationary environments. Evaluated on 200+ autonomous vehicles, FLAA-V2V sustains sub-300ms latency at 95% density and demonstrates 16.1% higher adversarial resilience than state-of-the-art FL baselines. This framework establishes a new paradigm for secure, adaptive federated learning in mission-critical ITS applications.