Abstract
The increasing prevalence of digital transactions and services has expanded the complexity and scale of fraudulent activity across a range of industries, including banking, e-commerce, health care, and insurance. Many fraud detection systems in these industries have traditionally employed federated models that require centralized data to determine and collect notable data points. Centralized systems can certainly be advantageous for situations like fraud detection, they create significant practical threats and challenges based on issues of privacy, security, compliance, and scale. Federated learning (FL) presents a novel paradigm, utilizing data from several institutions to develop collaborative machine learning models without transferring raw data. This enables the data to remain with the original host, maximizing privacy and complying with certain regulations, while enabling fraud detection that crosses organizations. This chapter provides an overview of FL and the basic principles behind it, an overview of FL architecture including client-server model, communications protocols, and model aggregation, and examples of fraud detection use cases from a wide range of sectors. The chapter will also examine the ways FL can break through information silos to detect complex and distrusted fraud schemes. We discuss important challenges such as data heterogeneity, as well as issues of communication overhead, and adversarial issues. Lastly, we demonstrate some of the most up-to-date solutions this recent development provides such as Differential Privacy (DP), Secure Aggregation, and Adversarial problem of robust anomalous detection systems. Finally, a brief discussion on additional evaluation measures that could be covered, such as false positive/false negative rates, privacy-based metrics, and efficiency of the systems. Montages of case studies in the practical deployment of FL to improve fraud detection applied in the financial services sector by banking consortiums, in healthcare by insurers, and in e-commerce by a growing collection of companies. The discussion of federated fraud detection concluded with alternatives for future work such as real-time federated analytics, syndication across different verticals, and developing intelligence on consumer devices (e.g., smartphones). In conclusion, FL is not only an enhancement in terms of capabilities as it relates to being a type of collaborative intelligence, but it provides for the imperative need for data privacy and resilient operations. As FL technology matures it will provide the foundational level of trustworthiness and privacy-awareness needed for the next generation of fraud avoidance systems.