Network intrusion detection systems (NIDSs) based on machine learning have been attracting much attention for its potential ability to detect unknown attacks that are hard for signature-based NIDSs to detect. However, acquisition of a large amount of labeled data that general supervised learning methods need is prohibitively expensive, and this results in making it hard for learning-based NIDS to become widespread in practical use. In this paper, we tackle this issue by introducing semi-supervised learning, and propose a novel detection method that is realized by means of classification with the latent variable, which represents the causes underlying the traffic we observe. Our proposed model is based on Variational Auto-Encoder, unsupervised deep neural network, and its strength is a scalability to the amount of training data. We demonstrate that our proposed method can make the detection accuracy of attack dramatically improve by simply increasing the amount of unlabeled data, and, in terms of the false negative rate, it outperforms the previous work based on semi-supervised learning method, Laplacian regularized least squares which has cubic complexity in the number of training data records and is too inefficient to leverage a huge amount of unlabeled data.
展开▼
