Customs fraud detection using a gradient boosting approach for joint classification and risk estimation

Customs revenue represents a major source of funding for most governments. Accordingly, customs administrations worldwide have promoted advanced analytics techniques to detect fraud and enhance revenue collection. However, inspecting imported commodities and auditing tax declarations is time-consuming, costly, and dependent on the inspector's expertise. Consequently, machine learning-based solutions have become critical for identifying fraud and minimizing revenue at risk. This paper applies the Dual-learning XGBoost-Based Approach (DXGBA) to the customs fraud detection domain. It demonstrates its ability to jointly detect fraud and estimate the corresponding revenue impact within a single boosting framework. The problem of under-valued imports is formulated as a dual supervised learning task, where DXGBA performs simultaneous classification and regression via a joint objective. The model ranks declarations based on their predicted revenue risk. Based on this ranking, it supports decision-makers in prioritizing inspections and enables maximizing revenue recovery under limited auditing capacity. To alleviate class imbalance, resampling strategies including the Synthetic Minority Oversampling Technique (SMOTE) and Random Undersampling (RU) were investigated. DXGBA classifies declarations as "fraudulent" or "non-fraudulent" and estimates the revenue at risk. Comparative experiments on a benchmark customs dataset show the high performance of DXGBA over single-task and state-of-the-art baselines. Results indicate that DXGBA recovers up to 87.98% of revenue by auditing only 10% of declarations. Furthermore, two enhancement pipelines, including tree-based embeddings and autoencoder-based deep feature representations, were examined. They both yield further gains in accuracy and revenue estimation performance.