Tensor-Empowered Federated Learning for Cyber-Physical-Social Computing and Communication Systems
Laurence T. Yang, Ruonan Zhao, Debin Liu, Wanli Lu, Xianjun Deng
Abstract
The deep fusion of human-centered Cyber-Physical-Social Systems (CPSSs) has attracted widespread attention worldwide and big data as the blood of CPSSs could lay a solid data cornerstone for providing more proactive and accurate wisdom services. However, due to concerns about data privacy and security, traditional data centralized learning paradigm is no longer suitable. Federated Learning (FL) as an emerging distributed privacy-preserving machine learning paradigm would have great research significance and application values. Although few survey papers on FL already exist in the literature, the survey about FL from the perspective of human-centered CPSSs and tensor theory is lacking. Toward this end, we first introduce the CPSSs and deeply analyze the correlations among humans, cyber space, physical space and social space, as well as the opportunities brought by it. Afterwards, we present an overview of FL and then review extensive researches on FL in terms of resources management, communication, security and privacy protection, which provides a shortcut for readers to quickly understand and learn FL. Furthermore, the theory about tensor representation, operation and decomposition for handling massive, multi-source heterogeneous big data and corresponding applications are described. By leveraging the advantages of tensor in unified modeling, dimensionality reduction, and feature extraction, a framework and three tensor-empowered solutions are provided to solve these challenges about heterogeneous resource management, communication overhead together with security and privacy. Finally, some future research directions are listed for looking forward to inspiring more readers to devote themselves to researching tensor-empowered FL for human-centered CPSSs in the future.