Secure Privacy Preserving Banking Customer Churn Prediction Using Federated Learning and Fully Homomorphic Encryption

Abstract

Banking domain is interested in customer churn prediction applications due to the rising competition with financial technologies (FinTech). This fierce competition is impacting banks market share, and it was found that it’s much easier and less costly to keep existing customer rather than acquiring new customers to the bank. Secure privacy preserving Customer Churn Prediction is a challenging and interesting area for research. Federated Machine Learning (FedML) has been proposed to resolve privacy problem, by using federated learning (FL) to apply Machine Learning (ML) prediction at banks locally and was proven to be one of the most effective solutions for this challenge. However, some gaps are identified for using federated machine learning (FedML) like the security attacks targeting the aggregation server or communication with the clients. Accordingly, this research proposes securing FedML vulnerabilities using Fully Homomorphic Encryption (FHE) encryption through a secure privacy preserving framework for customer churn prediction. The proposed framework guarantees the privacy preserving of customer data using Federated Machine Learning (FedML) while securing the aggregation and communication against vulnerabilities by a (FHE) provably secure algorithm. The proposed solution is demonstrated using a public dataset to predict the customer churn of 3 bank clients in different locations. FedML is applied to ensure data privacy for each client by training the model locally while only sharing the updates. FHE is used to encrypt all the updates, model aggregation and model prediction. Prediction accuracy is compared for the global model, the FedML without encryption and the FedML with FHE encryption using neural network binary classifiers. The proposed framework achieved high prediction accuracy, very close to the baseline, in addition to providing privacy and security safeguards that are mandated in banking domain.