所选语言本页无任何内容 Chinese : 中文 因此返回到 English : English

WCO BACUDA experts develop and share a neural network model to assist Customs to detect potential fraudulent transactions

19 五月 2020

WCO BACUDA experts develop and share a neural network model to assist Customs to detect potential fraudulent transactions

As part of the WCO BACUDA (Band of Customs Data Analysts) project with the Institute of Basic Science (IBS) and the National Cheng Kung University (NCKU), WCO has developed a Dual-Attentive-Tree-aware-Embedded (DATE) neural network model to assist Customs administration to better detect transactions presenting risks of fraud. The DATE model has been accepted by KDD2020[1] Conference (Applied Data Science Track) and will be published in the KDD2020 proceedings as a full paper[2].

The WCO BACUDA project was launched in September 2019 as a collaborative research platform focused on data analytics. With the participation of Nigeria Customs Service (NCS), BACUDA experts successfully developed the DATE model, and have been implementing a pilot test to verify its performance with real-time import data of the two Nigerian ports in Tin Can (in Lagos) and Onne (in Port-Harcourt) since March 2020.

The model employed a cutting-edge Artificial Intelligence (AI) mechanism called “ATTENTION” that is used as a language translation tool and for self-driving cars. Thanks to this innovative technology, the model has outperformed other traditional machine learning models (such as XGBoost) in detecting potential fraudulent transactions. The model noticeably outperforms even with relatively small-sized training data (from countries with low trade volumes) and low inspection rates (from countries with huge trade volumes).

How does the DATE model works?

  • Imagine that you are the head of a Customs Targeting Centre (neural network) composed of 100 risk analysts (decision trees). You want the analysts to report the probability of undervaluation and estimate the additional revenue from the inspection (dual-task).
  • Analyzing 100 different reports in order to make a final decision is a tedious task. Averaging the different predictions will lead to a loss of valuable information that may be hidden in any of the 100 reports. That is where the DATE model becomes handy as it keeps all the information while focusing on more important data. Some of the advantages of the model are:
    • If there are a majority group of reports significantly similar to each other, it allows you to pay more attention to those reports;
    • If you have analysts specialized in specific HS codes and importers of your target (i.e., the given import), you may pay more attention to their reports; and
    • Your final decision will reflect the reports that attract higher attention.

Interested in using the DATE model?

The DATE model is open-source and you can download it from: https://github.com/Roytsai27/Dual-Attentive-Tree-aware-Embedding.

You may extract transaction-level import data from your Customs clearance system and input them into the DATE model. It will provide you, for a given import, the risk of undervaluation and an estimate of additional revenue if inspected. 

Technical support

WCO BACUDA experts are developing a user guide/manual of the DATE model, available soon from the WCO webpage. For further customized support, WCO invites Members to contact WCO research unit (Research@wcoomd.org) to organize a joint test of the DATE model with BACUDA experts.



[2] *Sundong Kim, *Yu-Che Tsai, Karandeep Singh, Yeonsoo Choi, Etim Ibok, Cheng-Te Li, and Meeyoung Cha. DATE: Dual Attentive Tree-aware Embedding for Customs Fraud Detection. To appear in Proceedings of ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD). San Diego, CA, USA, August 23-27, 2020. (*: equally-contributed first authors)