Anomaly Dataset Augmentation Using the Sequence Generative Models

Abstract

In cyberspace, anomalies including intentional attacks grow up in their size and diversity. Although using the Intrusion Detection System (IDS) as a solution is helpful to some degree, there is an unsolved problem; the low performance of IDS due to lack of enough attack data. Recent approaches to solving this problem use an unsupervised deep learning-based technique called Generative Adversarial Networks (GANs). Because GAN variants show great performance in image augmentation, some research tries to apply GANs to cyberspace by domain conversion from binary to image. However, the attribute of cyberspace benchmarks is different from that of images. In this paper, we propose using sequence-based generative models such as Sequence Generative Adversarial Nets (SeqGAN) and Sequence to Sequence (Seq2Seq) to augment the ADFA-LD dataset, a sequence call based benchmark. Experimental results show that the performance is better when training ADFA-LD with augmented data from SeqGAN and Seq2Seq than training only the original dataset.