Optimizing Efficient Personalized Federated Learning with Hypernetworks at Edge

The recent advances in 5G and mobile edge computing facilitate the rapid development of the Internet of Things (IoT), enabling collective intelligence with data support from a massive number of IoT devices. Meanwhile, federated learning (FL) has emerged as a promising solution for collaborative training while preserving user privacy, which, however, is prone to poor learning performance in large-scale IoT scenarios. On the one hand, due to the task heterogeneity in various IoT scenarios and data heterogeneity (non-IID) across different IoT clients, more than traditional FL models derived from a Uniform FL (UFL) learning architecture are required to satisfy the diversified demands of each client. On the other hand, with the proliferation of IoT devices, achieving low latency and low communication costs with traditional FL architectures becomes even more challenging. In this article, we argue that customizing a specific model for each client is an urgent requirement, calling for the development of UFL to Personalized FL (PFL). In addition, the confluence of PFL and edge computing further provides opportunities for practical implementation in the 5G IoT environment. Accordingly, we propose EdgeFHN, an edge computing-based personalized federated learning framework, which can strike a balance between collaborative FL training among different clients and efficient model personalization for each client. Experiments on image classification demonstrate the superiority of our framework in improving accuracy and reducing communication overhead compared with other state-of-the-art solutions.