Federated Learning Enhanced YOLOv8 for Privacy-Preserving Student Classroom Behavior Recognition

Abstract

<title>Abstract</title> In the field of smart education, real-time object detection for analyzing student classroom behaviors provides valuable, objective data to help teachers optimize teaching methods and improve student learning experiences. This supports a positive and engaging classroom environment. However, models like YOLOv8 face challenges in real-world settings, such as varying object scales ("far small near large"), frequent occlusions, class imbalances, and privacy concerns when sharing data across institutions. Current datasets are often simulated, small in scale, and lack diversity, which limits their ability to reflect actual classroom conditions. To address these issues, this paper introduces FedYOLO-Behavior, a federated learning (FL) enhanced YOLOv8 framework that ensures privacy while recognizing behaviors effectively. We build a large, real-world database covering educational stages from kindergarten to university, combined with open-source data for augmentation. Local models are improved with multi-head self-attention (MHSA) for better context understanding, Ghost Convolution for efficiency, and Focal-EIoU loss to handle imbalances and small objects. FL with differential privacy allows safe collaboration between schools without sharing raw data. Experiments show significant improvements: mAP from 81.8% to 85.6%, precision from 77.9% to 82.3%, recall from 75.4% to 79.1%, inference speed increased by 18.2% (reaching 112 FPS), and parameters reduced by 25.1%, with a privacy budget of ε=0.9. This work promotes innovative, secure AI applications in education, contributing to national goals in technology and harmonious learning environments.

Affiliated Institutions

• Nanjing Xiaozhuang University CN

Related Publications