Embracing the Crowd: Robust Federated Learning for Edge Intelligence

With the rapid growth of edge intelligence, federated learning technologies have received a great deal of attention. Compared with traditional centralized machine learning based on cloud computing, federated learning in edge environment uses mobile edge devices to cooperatively train machine learning models. However, computational ability and transmission conditions are also varied among FL participants. Thus, robustness is one of the key issues for FL in edge scenarios. This article aims to focus on robust federated learning from the following four aspects: resilience to data variability and distribution, fault tolerance, preserving training performance against various attacks and proactive audit function. In this article, we provide an overview of the development history of federated learning for edge intelligence, and we gradually introduce it from three aspects: distributed machine learning, federated learning, and distributed federated learning. Then, in view of the threat problem of robust federated learning, relevant work is studied from three aspects: noisy data, noisy communication, and targeted and untargeted attacks. Then, relevant research is conducted on the building robust and reliable FL on the edge, focusing on three aspects: robust estimation, benefits against targeted and untargeted attacks, and cryptographic audits. Finally, we discuss the future directions and open problems for federated learning.