Deep Learning-Based International Trade Fraud Detection System

Abstract

Fraudulent activities in international trade cause significant economic losses and disorder. Traditional rule-based and statistical modeling methods struggle to cope with increasingly sophisticated fraud tactics. This paper proposes a deep learning-based model for identifying international trade fraud, integrating mechanisms such as embeddings, convolution, and attention to automatically extract features from high-dimensional trade data and capture hidden fraud patterns. Tested on a large dataset comprising hundreds of millions of real records, the model demonstrates outstanding performance with accuracy of 99.21%, F1 score of 91.72%, and AUC of 98.65%, significantly outperforming traditional methods. Furthermore, comprehensive evaluations confirm its robustness, showing controlled performance degradation under data poisoning and adversarial perturbation attacks, highlighting its excellent defensive capabilities. This work provides substantial technological support for anti-fraud efforts, crucial for maintaining international trade order and promoting sustainable economic development.