Research on Improving the Robustness of Tax Fraud Detection Model Based on Adversarial Attack and Defense

In view of the adversarial attack risks faced by tax fraud detection models in real-world applications, this paper takes the US tax environment as the background to construct a robustness improvement system that integrates attack generation and defense optimization. The study first completes the modeling foundation based on IRS SOI data and synthetic fraud datasets, and then uses FGSM and PGD to generate controllable perturbation samples to simulate attack scenarios. In risk assessment, the degree of model performance degradation is quantified by multiple indicators. In order to improve the robustness of the model, defense mechanisms including adversarial training, regularization and integration strategies are designed, and their effectiveness is verified by comparing t-SNE and multi-round AUC. Experimental results show that the systematic framework proposed in this paper can effectively suppress the impact of attacks and provide technical support for intelligent auditing of tax fraud.