rFedKD: A Reverse Federated Knowledge Distillation Method for Communication Efficiency

Artificial intelligence (AI) rapidly advances technological innovation, particularly in data processing and intelligent decision-making. Edge computing (EC) addresses key challenges in AI deployment, such as reducing latency and reliance on centralized cloud infrastructure by enabling processing at the network edge. However, EC faces limitations in managing computational complexity, latency, and resource constraints on edge devices. To overcome these challenges, we propose a novel reverse Federated Knowledge Distillation (rFedKD) method. Unlike traditional knowledge distillation, rFedKD extracts knowledge from small, personalized models at the edge and integrates it into a large central model. This approach aggregates diverse information while maintaining client-specific knowledge. The central model, leveraging its generalization capabilities, improves the accuracy of personalized edge models and accelerates training processes.The experimental results demonstrate the effectiveness of rFedKD, achieving a 15% improvement in accuracy and reducing communication rounds by 20 compared to the latest methods. This enhances system efficiency and user experience, establishing rFedKD as a promising solution for advancing edge AI.