A Privacy-Preserving Approach to Health Insurance Fraud Detection Using Vertical Federated Learning

In fraud detection, centralized approaches often face challenges related to data protection, security, and potential data breaches. Such methods require sensitive healthcare and insurance data to be pooled in one location, which increases vulnerability to misuse. This paper introduces FraudNetX, a privacy-preserving fraud detection framework, by utilizing Vertical Federated Learning (VFL) to address centralized system limitations. VFL enables models to be trained collaboratively while ensuring data privacy and security through quantifiable Differential Privacy (DP) guarantees (ε = 1.0, δ = 1 × 10⁵). FraudNetX implements a noise injection based on Differential Privacy (DP) with Gaussian noise (s = 1.2) in the process of training the model to guarantee confidentiality of the personal data. This research entails two partner organizations, which are a hospital and an insurance company, in an actual VFL configuration. The model is trained on 10 communication rounds in this federated setup, and the local optimization of each client is followed by the global aggregation step. Hospitals and insurers can learn fraud patterns without sharing their data. The proposed FraudNetX is a hybrid architecture which is composed of Feedforward Neural Networks (FFNNs) and transformer encoders. An adaptive weighting strategy has been applied to handle category imbalance concern and enhance recall of a few categories which are hard to detect, especially in fraud involving minorities, balancing the recall performance. The framework also includes a decision model that uses hospital data and claim behavior to classify each claim as legitimate, under review, or fraudulent. The experimental evaluation on the real-world dataset demonstrates that FraudNetX enhances the accuracy and F1-score of fraud detection (accuracy = 99.91%, F1 = 99.94%, ROC-AUC = 0.98) but does not violate data privacy.