Print Email Facebook Twitter Customs Risk Assessment Based on Unsupervised Anomaly Detection Using Autoencoders Title Customs Risk Assessment Based on Unsupervised Anomaly Detection Using Autoencoders Author Oosterman, D.T. Langenkamp, W.H. van Bergen, E.L. Contributor Arai, K. (editor) Publication year 2022 Abstract In this paper we describe our initial findings on a method for improving anomaly detection on a dataset with scarcely labeled data, based on an ongoing use-case with the Belgian Customs Administration (BCA). Data on shipping containers is used to predict the level of risk associated with a shipment, as well as the probability that the shipment is fraudulent. The absence of labeled data prevents the use of super vised machine learning techniques and calls for unsupervised analysis. We employ an autoencoder to learn the distribution of the dataset and detect anomalies, under the assumption that only a fraction of all shipments is fraudulent. The absence of labels in the dataset complicates the evaluation of the autoencoder’s performance. A qualitative approach is taken to assess the assess the properties of the detected anomalies. The variable distributions of the anomalies differ significantly from variable distributions in the complete dataset and are marked interesting by domain experts. To obtain an impression of the quantitative performance in the absence of ground-truths, synthetic data is generated using a variational autoencoder. The preliminary qualitative and quantitative results suggest that autoencoders can provide value for customs risk assessment. Subject Customs risk assessmentUnsupervised learningAnomaly detectionAutoencoderVariational autoencoderSynthetic data genera To reference this document use: http://resolver.tudelft.nl/uuid:9de8d130-af78-4693-a4ab-d77f3e3cb088 TNO identifier 962742 Publisher Springer Science and Business Media Deutschland GmbH ISBN 9783030821920 ISSN 2367-3370 Source Lecture Notes in Networks and Systems, Intelligent Systems Conference, IntelliSys 2021, 2 September 2021 through 3 September 2021, 668-681 Document type conference paper Files To receive the publication files, please send an e-mail request to TNO Library.