Vision-Based Fraud Detection in Insurance Claims

Abstract

With recent developments in information technology and artificial intelligence (AI), image-based insurance claims forgeries are now easier than ever to produce. Fundamental to an effective detection of such fraudulent insurance claims is the development of AI-models for automatic processing of claim-specific image data for the purpose of forgery detection. Although general forgery detection models are being developed, and machine learning methods are becoming increasingly popular in insurance fraud detection, practical applications of computer vision techniques to claim-specific image data are sparse. Here, we adapt the recently proposed Hierarchical Fine-Grained Image Forgery Detector and Localizer (HiFi-Net), for practical application in insurance image forgery detection and show that it can accurately detect image forgeries in natural images from the real-world car insurance claims domain. Furthermore, we report that the integration of a Vision-Transformer (ViT)-based masked auto-encoder into the network reduces forgery detection performance, due to the difference in pre training objectives. Application of our network to a document-type forgery detection task reveals cross-domain generalizability. The addition of a textual feature extractor slightly increased model performance, highlighting the strength of hybrid vision-textual architectures. However, it struggles to detect real-world document-type manipulations, mainly because of domain shifts and differences in forgery attributes between datasets. Lastly, we construct a hierarchical car-based insurance image forgery detection dataset to facilitate our study. This study provides us with more insight into forgery detection networks and how a single model can detect various types of image forgeries. Moreover, it highlights the importance of targeted fine-tuning in addressing domain-specific challenges, ensuring that AI-driven forgery detection remains effective and reliable for real-world insurance ap plications.