Privacy Preserving Federated Object Detection with Selective Semantic Fusion

Federated learning for medical image analysis faces challenges including data heterogeneity, privacy constraints, and communication overhead. This paper proposes FedSelect-YOLO, a novel federated object detection framework that employs selective semantic token transmission and dynamic multimodal feature selection to improve detection accuracy and communication efficiency. Evaluated on LIDC-IDRI, INbreast, and ISIC+Derm7pt datasets, FedSelect-YOLO achieves up to 87.1% mAP, 87.6% F1-score, and 84.4% IoU, outperforming baseline FedAvg-YOLO by over 6.7% mAP and reducing communication overhead by 80%. Under differential privacy with ε = 0.1, the model maintains robust performance with less than 2% metric degradation. Interpretability analysis using Grad-CAM and attention visualization confirms modality-aware feature attribution aligned with clinical relevance. These results demonstrate FedSelect-YOLO’s potential for efficient, privacy-preserving, and interpretable federated medical image classification in real-world hospital networks.