A Lightweight Splitted Federated Learning Algorithm for Financial Crime Detection

Federated learning (FL) has become a critical solution for privacy-preserving machine learning in the financial sector, allowing institutions to collaborate on model training without sharing sensitive data. As financial data is highly regulated, FL ensures that raw data remains localized, mitigating risks of data breaches and compliance violations. However, traditional FL faces challenges posed by limited client hardware capabilities and communication security. To mitigate these challenges, we propose a more versatile and secure FL framework. To maximize participation from various clients, we introduce the concept of split learning, which involves dividing the model into several parts so that clients with poorer hardware can also contribute. Furthermore, we utilize a combination of symmetric and asymmetric encryption algorithms to provide dual security for the communication process between clients and the server. Extensive experiments have been conducted to demonstrate the feasibility and effectiveness of our proposed method.