A unified federated learning framework for wireless communications: Towards privacy, efficiency, and security

Training high-quality machine learning models on distributed systems is a critical issue to achieve edge intelligence in wireless communications. Conventional data-driven machine learning approaches are infeasible due to non-IID data caused by privacy issues and the limited communication resources in wireless networks. Besides, considering the complex user identities, the training process also faces the challenges of Byzantine devices, which can inject poisoning information into models. In this paper, we propose a two-step federated learning framework, robust federated augmentation and distillation (RFA-RFD), to enable privacy-preserving, communication-efficient, and Byzantine-tolerant on-device machine learning in wireless communications. RFA is a method to tackle the problem of non-IID local data, which firstly trains local data generators on edge devices, then trains a global generator in the cloud server according to the IID dataset generated by the uploaded local generators, and finally, devices rectify non-IID dataset by downloading the global generator. After obtaining IID local data in edge devices, RFD is implemented to improve the performance of local models, in which devices only share the local information of models' outputs to reduce communication overhead. By employing a detection and discard mechanism in both RFA and RFD, our framework achieves robustness to the influence of Byzantine devices. Experiments show the effectiveness of RFA-RFD on preserving privacy, correcting non-IID data, reducing communication overhead, and resisting Byzantine devices, without much loss of accuracy compared with existing state-of-the-art methods