FedDroneQ: A Quantized Federated YOLOv5 Framework for Efficient UAV Object Detection

This paper presents FedDroneQ, a reproducible federated learning framework that adapts the YOLOv5 detector for privacy-preserving and communication-efficient UAV object detection. FedDroneQ trains lightweight YOLOv5s models on distributed VisDrone subsets and compresses client-to-server updates using 8-bit symmetric quantization. To mitigate statistical heterogeneity across UAV clients, the framework supports Dirichlet non-IID partitioning and incorporates a FedProx-style proximal regularizer. Beyond the core algorithm, FedDroneQ writes auditable artifacts to disk (configuration, per-round communication accounting, and update-norm summaries) and exposes diagnostic probes for localization and prediction behavior. Our experiments emphasize reproducible end-to-end validation, including CUDA runs that yield non-zero training/evaluation losses and non-zero update norms, confirming that federated rounds perform real parameter updates. We also report a small-subset detectionaccuracy sanity check using TorchMetrics mAP on a capped 200-image evaluation.