Autonomous Vehicles (AVs) increasingly depend on artificial intelligence (AI) for traffic sign recognition; however, adversarial vulnerability, privacy-preserving training, and real-time deployment constraints remain significant challenges. This paper presents FALCONS, a federated learning (FL)-based framework for traffic sign classification in Vehicular Ad Hoc Networks (VANETs), evaluated on the German Traffic Sign Recognition Benchmark (GTSRB). The proposed framework is lightweight, explainable, and operates in a serverless manner during testing and evaluation. A single-pass training paradigm on clean images is adopted, and robustness is assessed against multiple perturbations, including Snow Noise, Fast Gradient Sign Method (FGSM), and Red Patch Exploitation (RPE). Without adversarial retraining, the framework sustains >93–95% Top-25 accuracy under attack, while achieving 99.67% Top-1 and 100% Top-25 accuracy on clean data. To evaluate architectural generality, the same experimental pipeline is executed across ResNet-18, MobileNet-V2, and EfficientNet-B0 under identical conditions, enabling consistent and fair comparison. A decentralized file-based FL mechanism eliminates single-point dependency and supports offline, privacy-preserving collaboration among clients. Additionally, a Top-K robustness diagnostic is incorporated to analyze prediction stability under adversarial degradation in safety-critical scenarios. For deployment, ResNet-18 is optimized for low-latency inference and complemented by GUI-assisted inference and Excel-based logging for traceability. Experimental results averaged over five random seeds (mean ± standard deviation) ensure statistical reliability. Furthermore, a conceptual mapping between GTSRB and Indian MoRTH/IRC:67 standards highlights potential cross-domain applicability. Overall, the framework demonstrates strong robustness and computational efficiency within the evaluated experimental setting, while validation across heterogeneous datasets remains future work.
Patil et al. (Wed,) studied this question.