PrivRBFN: Building Privacy-Preserving Radial Basis Function Networks Based on Federated Learning

Federated learning is a promising solution for data owners who want to build machine learning models collaboratively. With the widespread use of radial basis function (RBF) networks in machine learning, it is appealing to use the technique of federated learning to build RBF networks on decentralized data, mainly when the data owners have restricted training data and computational resources. Although federated learning is privacy-friendly, the construction of RBF networks based on federated learning faces the risk of privacy leakage. In federated learning, the radial bases representing the training data distribution should be transmitted to the server to form the hidden neurons of a global RBF network. However, the transmission process unavoidably discloses data privacy and breaks the rule of federated learning, i.e., keeping data private at each site. In this paper, we propose a novel algorithm called PrivRBFN to build a privacy-preserving RBF network based on federated learning. PrivRBFN is tailored for RBF networks with Gaussian radial basis functions. It allows the clients involved in federated learning to upload approximate radial bases of their local models to the server for model aggregation, thereby preventing the leakage of data privacy. In addition, PrivRBFN can maintain the performance of the global RBF network. Experiments on publicly available datasets demonstrated that the classification accuracy of the model built by PrivRBFN is comparable to that of the non-privacy-preserving RBF network in federated learning.