Federated Learning for Telecom Fraud Detection: A Privacy-Preserving Approach to Overcoming Data Fragmentation and Enhancing Security

Over the last couple of years, the world has seen numerous frauds about telecom: fraud calls, phishing, and misutilization of personal numbers, among others. However, traditional methods of fraud detection cannot fit the shifting intricacy of these fraud schemes since they depend on data collection in a centralized way. While the arrival of deep learning improves the detection capability, it engenders considerable privacy risks and issues of data fragmentation. As one might guess, this work investigates federated learning as a decentralized solution to these limitations. The FL allows various organizations to train fraud detection models collaboratively while preserving data privacy via sharing only model updates, not raw data. This paper proposes a federated learning-based system for phone number fraud detection and defends personal data against various industries. We will discuss in more detail the advantages of federated learning in solving the "data island" problem and reducing the risk of privacy breaches in a distributed environment. The paper also looks at the use of horizontal and vertical federated learning in co-governance both within and across industries. Finally, we discuss the limitation of FL from a practical perspective by including problems arising from non-IID data and heterogeneity in systems and scalability. We also identify further works to be pursued in optimizing the performance of FL fraud detection with privacy preservation.