An Efficient Privacy-Aware Framework for Traffic Violation Detection Using Federated Learning

Traffic violations such as helmet non-compliance, triple riding, wrong-route driving, and mobile phone usage while riding are major contributors to road accidents. In this work, a privacy-preserving federated learning (FL) framework for traffic violation detection is proposed, leveraging Horizontal FL (HFL) with the Federated Averaging (FedAvg) algorithm. To ensure that raw video data from roadside CCTV cameras (edge devices) remains secure and decentralized, only model updates are transmitted to the central server. SSD MobileNetV2 is employed as the backbone for object detection, trained on the general traffic violation dataset (Roboflow variants) in COCO format. Experimental evaluation over 4 global communication rounds with 4 simulated clients demonstrates consistent performance improvements. Accuracy improved from 55.6% to 83.1%, while mean Average Precision (mAP) increased from 50.1% to 81.0%. Privacy leakage risk, measured via gradient leakage loss (MSE), decreased from 0.299 to 0.187, indicating stronger resilience against model inversion attacks. Communication overhead analysis revealed a linear increase from 40 MB in round 1 to 160 MB in round 4, consistent with model exchange requirements. The results validate that FL can achieve competitive detection performance while safeguarding sensitive surveillance data, making it a viable solution for real-world smart traffic management systems.