Fed-YOLOv10 framework for sustainable privacy-preserving real-time classroom behavior analysis

Monitoring student behavior in classrooms using computer vision raises severe privacy concerns, especially when videos must leave edge devices. We propose Fed-YOLOv10, the first real-time privacy-preserving federated object detection framework tailored for educational environments. By integrating detection-aware weighted aggregation (Det-FedAvg), dynamic structured pruning with RLE+gzip compression, strict differential privacy ( ε ≈1.0), and lightweight C3TR blocks with Dual Attention, Fed-YOLOv10 simultaneously achieves near-centralized accuracy (91.0% mAP@0.5), 8.2× lower communication overhead (9.8 MB/round), 62% reduced energy and carbon footprint, and 41% faster convergence under extreme non-IID conditions. Real-time inference exceeds 56 FPS on NVIDIA Jetson Orin Nano while raw video never leaves the edge. A one-month pilot deployment across Nanjing schools from primary to university classrooms confirms >94% agreement with manual observation and strong acceptance by teachers. Fed-YOLOv10 establishes a new benchmark for accurate, private, low-carbon, and practically deployable federated vision systems in real-world education scenarios.codes will be available at https://github.com/Pa55go/yolov10-Fed.git