I. Introduction
The digital age in which we currently live is wealthy with numerous kinds of data, which is generated by healthcare sensing devices, smart vehicles, industrial robots, and more [1]. In the meantime, this data is increasing day by day. As a result, recent years have seen a rapid development of data-driven machine learning (ML) technology, which can timely and intelligently extract useful knowledge from data to serve as the foundation for practical applications. ML has been extensively employed in smart clinical decision-making [2], smart assisted driving [3], intelligent manufacturing [4], and other real application fields. However, traditional ML typically involves centralizing scattered user data to one source, which conflicts with the growing awareness of data security and user privacy. In response, relevant laws exist to strengthen data security and privacy protection, such as General Data Protection Regulation. Moreover, this centralized training approach requires transmitting a large amount of raw data which brings high communication overhead.