Few-Shot Network Intrusion Detection Using Discriminative Representation Learning with Supervised Autoencoder

Abstract

Recently, intrusion detection methods based on supervised deep learning techniques (DL) have seen widespread adoption by the research community, as a result of advantages, such as the ability to learn useful feature representations from input data without excessive manual intervention. However, these techniques require large amounts of data to generalize well. Collecting a large-scale malicious sample is non-trivial, especially in the modern day with its constantly evolving landscape of cyber-threats. On the other hand, collecting a few-shot of malicious samples is more realistic in practical settings, as in cases such as zero-day attacks, where security agents are only able to intercept a limited number of such samples. Hence, intrusion detection methods based on few-shot learning is emerging as an alternative to conventional supervised learning approaches to simulate more realistic settings. Therefore, in this paper, we propose a novel method that leverages discriminative representation learning with a supervised autoencoder to achieve few-shot intrusion detection. Our approach is implemented in two stages: we first train a feature extractor model with known classes of malicious samples using a discriminative autoencoder, and then in the few-shot detection stage, we use the trained feature extractor model to fit a classifier with a few-shot examples of the novel attack class. We are able to achieve detection rates of 99.5% and 99.8% for both the CIC-IDS2017 and NSL-KDD datasets, respectively, using only 10 examples of an unseen attack.