Insurance Claims Fraud Detection via Federated Learning with Drift Correction and Self-Knowledge Distillation

Abstract

Introduction: Insurance claims fraud poses a critical challenge for insurers, leading to substantial financial losses and operational inefficiencies. While machine learning has shown significant promise in detecting fraudulent claims, its effectiveness is often constrained by stringent data protection regulations that create private data silos. These silos impede the free exchange of data across organizations, limiting centralized collaboration and hindering the development of robust, generalizable fraud detection models. Methods: To overcome these barriers, I propose pFed-ICKD, a novel federated learning–based algorithm tailored for insurance claims fraud detection. pFed-ICKD integrates multi-party insurance data while preserving data privacy through a federated framework. The method incorporates two key innovations: (1) a local drift correction mechanism that tracks and mitigates divergence between local and global model parameters using auxiliary variables, and (2) client-side self-knowledge distillation that retains task-specific knowledge under heterogeneous, non-IID data distributions—without incurring additional communication overhead. This design enables effective cross-insurer collaboration without requiring access to sensitive raw data. Results: Experimental evaluations demonstrate that pFed-ICKD achieves faster convergence and superior performance across diverse fraud detection tasks compared to existing federated baselines. Specifically, it converges 6.52 times faster than FedAvg implemented with a Keras Sequential architecture and achieves a precision of 93.5% in mixed settings with weak, non-IID client data distributions. Notably, it exhibits strong robustness under realistic conditions characterized by partial client participation and high data heterogeneity—common challenges in real-world insurance ecosystems. Discussion: A major contribution of this work is the joint optimization of personalization and stability in federated insurance claims fraud detection. Unlike conventional approaches that struggle with model drift under data heterogeneity, pFed-ICKD explicitly addresses this issue through local drift correction and self-knowledge distillation. By enabling secure, collaborative learning across fragmented data sources, my approach not only complies with data privacy regulations but also enhances detection reliability through richer, multi-dimensional insights—offering a practical pathway toward scalable and trustworthy fraud prevention in the insurance industry.