Privacy-Preserving Secure Framework for Intelligent Transportation Systems

Federated Learning (FL) is increasingly recognized as a promising approach that can be used to train various machine learning models across distributed clients while preserving the privacy of the training data. However, challenges remain when clients experience class deficits, where certain labels are missing from their local datasets. This study investigates the performance of object detection models under such conditions, focusing on YOLOv3 and Tiny-YOLOv3. We show that YOLOv3 achieves significant improvements in Mean Average Precision (mAP), surpassing the baseline by at least 10%. The highest mAP of 79.24% was achieved despite each client lacking one class label. Moreover, we observed that clients began detecting missing labels at varying communication rounds, illustrating FL's adaptability to real-world scenarios. In contrast, lightweight models such as Tiny-YOLOv3 showed significant mAP reductions, highlighting their limitations in FL setups.