Privacy-Aware Federated Learning for Intelligent XR Systems at the Wireless Edge

Mixed Reality (XR) systems increasingly integrate artificial intelligence to enable real-time perception, classification, and interaction across healthcare, education, and industrial domains. However, these systems often require transmitting highly sensitive data, such as body movements, biometric signals, and spatial environments, to centralized servers for training, raising serious privacy and bandwidth concerns. In this paper, we present a novel federated learning (FL) framework for intelligent XR systems that performs on-device training and model aggregation at the wireless edge. Our architecture enables multi-user XR devices, including headsets, wearables, and IoT sensors, to collaboratively learn user behavior patterns without exposing raw data. We evaluate our system across simulated XR datasets involving gesture recognition and avatar personalization, comparing centralized and federated configurations in terms of model accuracy, communication cost, latency, and privacy leakage. Results show that our FL framework achieves comparable performance to centralized training while significantly reducing bandwidth consumption and exposure to inference attacks. We also present a detailed analysis of inference time and convergence behavior in edge environments. This work demonstrates the feasibility and necessity of privacy-preserving learning for next-generation XR networks and lays the foundation for future deployment in sensitive domains such as telehealth, rehabilitation, and smart training.