Introduction
In international trade, customs agencies manage vast quantities of data generated by import declarations. These documents contain critical information regarding cross-border shipments, including item descriptions, valuations, and participant identities. While this data is essential for identifying smuggling and tax evasion, its sensitive nature creates a "data silo," preventing the broader research community from developing and benchmarking advanced fraud detection algorithms.
To address this accessibility gap, researchers from KAIST, GIST, and the Korea Customs Service (KCS) have introduced a high-fidelity, privacy-preserving synthetic dataset. This benchmark allows for the open development of machine learning models without compromising the legal or security requirements of actual customs records.
Methodology
The researchers utilized a Conditional Tabular GAN (CTGAN) to generate 54,000 synthetic import declarations. CTGAN was chosen for its ability to handle the multi-modal nature of tabular data, specifically imbalanced discrete variables and complex distributions.
The generation process involved three key phases:
- Preprocessing and Anonymization: Identifiable information was converted to IDs, and Gaussian noise was applied to price data to mitigate re-identification risks.
- Correlation Maintenance: To ensure logical consistency (e.g., matching product codes with appropriate tax rates), the authors employed column aggregation and post-processing. Item prices were reconstructed mathematically from net mass and unit prices to preserve internal relationships.
- Validation: The synthetic data was evaluated using the SDMetrics library, focusing on column shape similarity, pair-wise trends, and the synthesis of unique rows to ensure the model did not simply memorize the training data.
The Customs Import Declaration Dataset
The resulting dataset comprises 54,000 artificially generated trades based on 24.7 million real declarations from the Korea Customs Service (2020–2021). It includes 22 key features such as Office ID, HS6 product code, Country of Origin, and Net Mass. Crucially, the dataset includes binary labels for general fraud and ternary labels for "Critical Fraud," which represents threats to public safety.
Evaluation and Results
The synthetic dataset was benchmarked against real-world data to ensure "trend preservation." The authors found that model performance rankings remained consistent across both datasets. For instance, Gradient Boosted Decision Trees (GBDTs) like LightGBM significantly outperformed simpler models like Logistic Regression on both real and synthetic data.
Key performance metrics included:
- Statistical Fidelity: High similarity scores were achieved in Column Pair Trends (0.9569) and Boundary Adherence (0.9869).
- Predictive Accuracy: LightGBM achieved a precision of 0.7783 at a 5% inspection rate on the synthetic data.
- Feature Consistency: SHAP value analysis confirmed that models prioritized the same features (Importer ID, Item Price, HS6 Code) regardless of the data source.
Conclusion and Future Work
This synthetic benchmark represents a significant step toward collaborative AI research in customs fraud detection. While the current dataset assumes fully labeled data, future work will explore "label scarcity" scenarios and the use of Diffusion Models for even higher-quality synthesis. By providing a realistic, open-access alternative to sensitive government records, this work facilitates the global development of more effective trade security measures.
Find the Data: The authors have made the dataset and baseline codes available at https://bit.ly/customs-dataset.
Summary updated