Anomaly Detection Based on Discriminative Generative Adversarial Network

Benjamin Appiah ； Zhiguang Qin ； Obed Tettey Nartey ； Brighter Agemang ； Ansuura JohnBosco Aristotle Kanpogninge

《International Journal of Network Security》 23卷4期 (2021/07) Pp. 718-724

https://doi.org/10.6633/IJNS.202107_23(4).18

摘要

Generative adversarial network (GAN) has recently achieved remarkable success in anomaly detection. The GAN structure consists of a generator as an attacker and a discriminator as a classifier, all competing against each other. Most of the available classifiers are trained on softmax cross-entropy (SCE) loss function. However, the SCE loss is less discriminative in classifying malicious samples. To improve the discriminative power of these GAN's classifiers, this paper proposes a discriminative GAN, D-GAN, for short. The key idea is to replace the SCE loss with Mahalanobis distance loss to induce strong intra-class compactness to construct high-density regions, which are essential to discriminating against new unseen classes. We evaluate D-GAN on MNIST and KDD99 datasets. The results demonstrate that D-GAN can achieve superior performance compared with several state-of-the-art SCE loss learning-based defending methods.

關鍵字

Anomaly Detection ； Distance Metric Learning ； Generative Adversarial Network