OMASGAN: Out-of-distribution Minimum Anomaly Score GAN for Anomaly Detection

Abstract

Generative models trained in an unsupervised manner may set high likelihood and low reconstruction loss to Out-of-Distribution (OoD) samples. This leads to failures to detect anomalies, overall decreasing Anomaly Detection (AD) performance. In addition, AD models underperform due to the rarity of anomalies. To address these limitations, we develop the OoD Minimum Anomaly Score GAN (OMASGAN) which performs retraining by including the proposed minimum-anomalyscore OoD samples. These OoD samples are generated on the boundary of the support of the normal class data distribution in a proposed self-supervised learning manner. Our OMASGAN retraining algorithm leads to more accurate estimation of the underlying data distribution including multimodal supports and also disconnected modes. For inference, for AD, we devise a discriminator which is trained with negative and positive samples either generated (negative or positive) or real (only positive). The evaluation of OMASGAN on image data using the leave-one-out method shows that it achieves an improvement of at least 0.24 and 0.07 points in AUROC on average on the MNIST and CIFAR-10 datasets, respectively, over other benchmark models for AD.

SLX: Similarity Learning for X-Ray Screening and Robust Automated Disassembled Object Detection

“…The general methodologies that could be applied to the X-ray security screening problem [2], [34] are: (a) OoD detection using deep generative models [2], [3], (b) Discriminative models with labelled data [6], [7], (c) Open-Set, or even Open-World, classification, combining OoD detection and discriminative models, and (d) Contrastive learning to achieve improved separation of the classes where similar image representations are attracted [18], [7], while different image latent feature representations are repelled and pulled apart [27], [18].…”

“…The samples h i , for i between 1 to N, are processed by the classification head, g(•), and the projection head, v(•), at the middle top and bottom, respectively. The outputs of g(•) are passed through the softmax [6], [22] to obtain the classes, 1 to K, [33], [36]. The labels are passed through the loss terms L 1 (x, y, F) and L 2 (F, v), in the loss in (1).…”

“…We denote the data by x where x i are the labelled image data with labels y i between 1 and K, where K is the number of classes. Let χ ∈ R d be the set of inputs, modelling objects of interest such as images [8], [9], and let L be the set of class labels. Let F : χ → L be a classifier [6], [7] that assigns a unique label from the set L to an element from χ .…”

“…Let χ ∈ R d be the set of inputs, modelling objects of interest such as images [8], [9], and let L be the set of class labels. Let F : χ → L be a classifier [6], [7] that assigns a unique label from the set L to an element from χ . Now, we denote the latent feature representation of our deep learning model by h, where h i = f (x i ) [18], [24].…”

“…The model g(•) is followed by a softmax output layer, i.e. the normalised exponential, softmax(l i ), to obtain the probability over the K classes [6], [25]. For ResNet, h i is the output after the final average pooling layer.…”