Effective and Robust Boundary-Based Outlier Detection Using Generative Adversarial Networks

Effective and Robust Boundary-Based Outlier Detection Using Generative Adversarial Networks

Outlier detection aims to identify samples that do not match the expected patterns or major distribution of the dataset. It has played an important role in many domains such as credit card fraud identification, network intrusion detection, medical image processing and so on. The inherent class imbalance in datasets is one of the major reasons why this problem is difficult to solve. The small number of outliers are not adequate to characterize their own overall distribution, which makes it difficult for classifiers to effectively learn the demarcation (boundary) between normal samples and outliers. To address this problem, we introduce an effective and robust Boundary-based Outlier Detection method using Generative Adversarial Networks (BOD-GAN). Here, we extract the border data containing normal samples and outliers, expand them to form the initial reference boundary outliers. With the min-max game between a generator and two discriminators in GAN, the boundary outliers are further augmented by BOD-GAN, which, together with the boundary normal data, provides the valuable demarcation information for classifier. However, the increase of the data dimension may bring some gaps in the initial boundary, which are difficult to effectively fill by the augmentation method alone. To address this, we innovatively add density-loss to the loss function of the generator to explore these boundary gaps, making our model rather robust even with the high dimensional data. The extensive experimental evaluation demonstrates that our proposed method has achieved significant improvements compared with existing classic and emerging (i.e., GAN-based) outlier detection methods.

Files associated with this item cannot be displayed due to copyright restrictions.