Generative models, such as Generative Adversarial Networks(GANs), have been used for unsupervised anomaly detection.While performance keeps improving, several limitations existparticularly attributed to difficulties at capturing multimodalsupports and to the ability to approximate the underlying dis-tribution closer to the tails, i.e. the boundary of the distribu-tion’s support. This paper proposes an approach that attemptsto alleviate such shortcomings. We propose an invertible-residual-network-based model, the Boundary of DistributionSupport Generator (BDSG). GANs generally do not guaran-tee the existence of a probability distribution and here, we usethe recently developed Invertible Residual Network (IResNet)and Residual Flow (ResFlow), for density estimation. Thesemodels have not yet been used for anomaly detection. Weleverage IResNet and ResFlow for Out-of-Distribution (OoD)sample detection and for sample generation on the bound-ary using a compound loss function that forces the samplesto lie on the boundary. The BDSG addresses non-convexsupport, disjoint components, and multimodal distributions.Results on synthetic data and data from multimodal distribu-tions, such as MNIST and CIFAR-10, demonstrate competi-tive performance compared to methods from the literature.
|Publication status||Accepted/In press - 25 Oct 2020|