Privacy Preserving Federated Knowledge Distillation for Distributed Intelligent Control in 6G Edge Networks

Abstract

6G edge networks are changing extremely quickly, and will require smart, real-time control of distributed information while keeping user data confidential. Centralized Learning methods will have adverse impacts due to additional processing, additional communication, and additional data exposure. This paper tackles these issues through the lens of distributed intelligent control in cyber-physical systems for 6G networks and proposes a framework called Privacy-Preserving Federated Knowledge Distillation (PP-FKD). The framework focuses on lossless data collaborative training through federated learning and knowledge distillation at disparate edge devices. For the first time, a multi-level privacy framework utilizing differential privacy and secure aggregation is proposed to maximize data protection. The proposed methods were experimentally evaluated in a 6G edge heterogeneous system, resulting in 27.4% less communication, 18.9% less time to converge the model, and 96.2% in control decision accuracy, compared to standard federated learning systems. In addition, control loops were faster than 21.7% in latency, allowing for near real-time control of intelligent systems. This is the first model to offer triple-layered protection to the privacy, efficiency, and performance of data in distributed systems. The model can be used in 6G and will be of tremendous use in autonomous systems, smart grids, and the industrial IoT. The developed system is consistent with the anticipations regarding the intelligent and adaptable technology within 6G networks, providing an efficient and secure approach to decentralized systems. Subsequent efforts will be directed toward the improvement of the energy consumption and the expansion of the system to an inter-domain federated ecosystem.