A Cloud-Native Federated Learning Architecture for Telecom Fraud Detection

We propose a cloud-native federated learning (FL) architecture, where a FL server is built with serverless microservices. We demonstrate this FL framework with a Telecom fraud detection use case, which trains a global model across multiple CSPs. Our results show that the global model trained by FL improves up to 23% in F1-score compared to locally trained models.

I. Introduction

Telecom fraud is the abuse of telecom products or services with the aim of illegally siphoning away money from either CSPs or its customers. According to a report by Europol [1], Telecom fraud is a fast-growing field of criminal activity costing CSPs US32.7 billion annually. Fighting Telecom fraud becomes particularly challenging with the arrival of 5G. First, 5G needs to support more end devices (10 times more than 4G LTE), which results in higher risk of fraud. Second, Telecom fraud detection needs information collected from 5G Core network functions. As 5G network functions move towards network edge, the complexity of fraud detection system increases.