on the effectiveness of generative adversarial network on anomaly detection

TL;DR

This paper proposes a novel unsupervised GAN-based model combining autoencoders and a new scoring function for anomaly detection, demonstrating superior performance on multiple benchmark and medical datasets.

Contribution

Introduces a new GAN-autoencoder hybrid model with a unique scoring function for improved anomaly detection accuracy.

Findings

Outperforms existing methods on SVHN, CIFAR10, MNIST, and leukemia datasets.

Achieves higher detection accuracy with slightly faster inference.

Effective in both general and medical anomaly detection tasks.

Abstract

Identifying anomalies refers to detecting samples that do not resemble the training data distribution. Many generative models have been used to find anomalies, and among them, generative adversarial network (GAN)-based approaches are currently very popular. GANs mainly rely on the rich contextual information of these models to identify the actual training distribution. Following this analogy, we suggested a new unsupervised model based on GANs --a combination of an autoencoder and a GAN. Further, a new scoring function was introduced to target anomalies where a linear combination of the internal representation of the discriminator and the generator's visual representation, plus the encoded representation of the autoencoder, come together to define the proposed anomaly score. The model was further evaluated on benchmark datasets such as SVHN, CIFAR10, and MNIST, as well as a public medical dataset of leukemia images. In all the experiments, our model outperformed its existing counterparts while slightly improving the inference time.