Fig. 1

Overview of our proposed method and comparison with existing model-sharing federated learning. (a) Existing anomaly-detection framework for distributed journal-entry data based on model-sharing FL (adapted from18), which constructs an aggregated model without directly sharing raw data. (b) Our DC-based framework, which, in addition to not sharing raw data, exchanges only dimension-reduced intermediate representations, completes training in a single communication round, and allows model construction without connecting raw data to external networks.