A deep-learning–based antifraud system for car-insurance claims

The annual cost of vehicle insurance fraud is estimated to exceed 40 billion dollars. This is an enormous amount considering the number of new vehicles insured yearly. In terms of higher premiums, it implies that insurance fraud incurs an additional annual cost to each U.S. family of $400 to $700, on average. Many frauds can be attributed to previously reported damages, which are submitted a second time to the insurance company. In these cases, it does not suffice to check the customer’s history to identify them: Damaged car panels can be removed from one vehicle and reassembled on another to make an insurance claim on the second car. To deal with these fraud attempts, in this paper, we propose an end-to-end solution to support the special investigation unit of the insurance companies in their antifraud investigations. For each claim, we organize the images sent to the insurance company and analyze them to extract basic vehicle information. Subsequently, we use these images to identify any damage to the bodywork, and, finally, we verify that the damage has not already been processed in previous claims. To the best of our knowledge, this is the first published work that deals with this problem through a pipeline that covers the entire claim management process. To validate our proposal, we compare our solution with other state-of-the-art models for estimating image similarity. Our results show that our solution is, on average, superior by 15.27% in terms of mean average precision (mAP). In addition, we report the challenges faced in scaling such a system in a production environment. This aspect is often ignored, but applying these solutions to industrial settings is of fundamental importance. Our proposed end-to-end system can reduce by up to 18% the number of false positives produced by the damage reidentification system. We show that encapsulating several specialized components and merging their intermediate results leads to a 72% reduction in possible alerts. Finally, to support the discussion and comparison of explanations for this new task, we introduce a new dataset as a benchmark for damage reidentification.