A Study on Adaptive Resonance Theory-based Clustering Incorporating Federated Clustering with Local ε-Differential Privacy

Abstract

Federated learning is a machine learning method that merges and updates models obtained from multiple devices learning each dataset. Compared to conventional machine learning methods, federated learning is expected to significantly reduce the learning time by distributed computing on each device in parallel. In addition, federated learning combined with a data anonymization method, called differential privacy, can guarantee a high degree of privacy. In this paper, we propose two methods. First, in the case of federated learning with a clustering method called CIM-based ART (CA), which represent clusters by nodes, we propose a method of merging nodes by reapplying CA with the generated nodes as input data. Second, we propose a method that combines federated learning with CA and differential privacy as a method for learning while guaranteeing a high degree of privacy. Numerical experiments showed that federated learning with CA achieves faster learning while maintaining a similar degree of clustering performance compared to the conventional method of machine learning with CA. The clustering performance results for a real-world dataset trained with a method that combines federated learning with CA and differential privacy were used to investigate the effect of noise from differential privacy on clustering performance.