Research On Anti-Fraud Of Automobile Insurance Based On Federal Learning

As an important part of property insurance,motor vehicle insurance can effectively reduce the losses caused by various traffic accidents.According to the data disclosed by the CBRC,motor vehicle insurance accounted for 68% of the total income of property insurance premiums in 2022.Car insurance fraud cases are not uncommon.According to the statistical data of the intelligent risk control white paper of China’s insurance industry,car insurance fraud cases account for 7% of the total business volume of car insurance,and the fraud compensation amount accounts for at least 20% of the compensation amount.In 2022,China’s auto insurance claim settlement expenses totaled 505.575 billion yuan.According to this calculation,the insurance company’s auto insurance fraud leakage loss could reach 101 billion yuan.The serious increase in the operating costs of the insurance company makes the insurance company have to increase the premium,thus damaging the legitimate rights and interests of the insured and damaging the normal operating order of the automobile insurance market;On the other hand,successful fraud will have a negative "demonstration effect",which will erode the social integrity system.Therefore,increasing anti-fraud efforts in automobile insurance has an important positive effect on the policyholders,insurance institutions and the social integrity system.The performance of auto insurance anti-fraud algorithm requires a large amount of data for training.It is difficult to obtain a model with good detection performance based on the data of a single company.Therefore,it is necessary to expand the sample size and feature size,and conduct joint modeling with other insurance companies or other industry institutions,but this will inevitably lead to the risk of data privacy disclosure.This paper introduces the federated learning technology into the field of anti-fraud in automobile insurance.This technology can guarantee the problem of co-building models without leaving the local data,fully protect the information privacy of all participants,and improve the overall anti-fraud detection accuracy.Among them,horizontal federated learning is applicable to joint modeling between insurance companies,and vertical federated learning is applicable to joint modeling between insurance companies and banks,traffic management departments and other institutions.The first chapter of this paper is an introduction,which introduces the background and significance of this research,as well as the research status of scholars in this field,and introduces the article framework and innovation.The second chapter introduces the types,causes and current problems of anti-fraud detection of automobile insurance fraud,and points out the measures that insurance institutions should take in anti-fraud work.The third chapter introduces the concepts and definitions of each classification of federated learning,and expounds two commonly used information privacy transmission methods: homomorphic encryption technology and differential privacy technology,as well as a secure computing library based on python to realize differential privacy.In the construction of anti-fraud model,the fourth chapter will train multiple GBDT integrated learning models,such as Ada Boost,XGBoost,Cat Boost,Light GBM,etc.,and use stacking,blending,average,and other methods to fuse to obtain the optimization algorithm.Considering that according to the distributed machine learning environment of federated learning,the Oputuna technology is used to optimize parameter adjustment,and FLD-WA algorithm is used to solve the data skew problem under the federated learning scenario,It effectively solves the imbalance of fraud data.Chapter 5 introduces how to deploy federal learning in the anti-fraud practice of insurance companies,and how to build the experimental environment.The open source federal learning platform FATE developed by We Chat is used for the experiment.With the help of Kaggle and Tianchi Automobile Insurance Fraud Data Set,the EDA analysis and simulation of horizontal federal learning and vertical federal learning are carried out.The experiment shows that the performance of the model has improved significantly after the federal learning.Chapter 6 writes conclusions and improvements.This experiment focuses on supervised machine learning.For semi-supervised and unsupervised models,algorithms such as isolated forest,anomaly detection,LOF,etc.need to be used for experiments.