Client-Edge-Cloud Hierarchical Federated Learning Based on Generative Adversarial Networks

Federated learning technology is a key approach to ensuring data privacy protection and promoting data sharing across different domains. Traditional federated learning, which relies on cloud servers, suffers from high computation latency and low reliability. Edge-based federated learning can reduce communication latency but has limitations in terms of the number of accessible clients, making it unsuitable for large-scale data computation. To address the integration of the cloud, edges, and end devices in real-world scenarios, previous research has proposed hierarchical federated learning architectures. However, these approaches have drawbacks such as limited model training diversity, high training costs for end devices, and high communication costs. To overcome these challenges, we propose a client-edge-cloud hierarchical federated learning protocol based on Generative Adversarial Networks (GANs). By combining GAN technology, it mitigates the issues of client and edge server data heterogeneity, thereby enhancing the stability of model training. The protocol allows edge servers to synthesize virtual datasets through GANs, enabling separate training of the client-edge federated learning module and the edge-cloud federated learning module. We conducted experiments on the MNIST dataset, and the results indicate that the protocol outperforms traditional hierarchical federated learning protocols in multiple scenarios. The protocol offers greater flexibility in model training and reduces overall system training costs, particularly in terms of lowering client communication and computation expenses.