Generating practical adversarial examples against learning-based network intrusion detection systems

Abstract

There has been a significant development in the design of intrusion detection systems (IDS) by using deep learning (DL)/machine learning (ML) methods for detecting threats in a computer network. Unfortunately, these DL/ML-based IDS are vulnerable to adversarial examples, wherein a malicious data sample can be slightly perturbed to cause a misclassification by an IDS while retaining its malicious properties. Unlike image recognition domain, the network domain has certain constraints known as domain constraints which are multifarious interrelationships and dependencies between features. To be considered as practical and realizable, an adversary must ensure that the adversarial examples comply with domain constraints. Recently, generative models like GANs and VAEs have been extensively used for generating adversarial examples against IDS. However, majority of these techniques generate adversarial examples which do not satisfy all domain constraints. Also, current generative methods lack explicit restrictions on the amount of perturbation which a malicious data sample undergoes during the crafting of adversarial examples, leading to the potential generation of invalid data samples. To address these limitations, a solution is presented in this work which utilize a variational autoencoder to generate adversarial examples that not only result in misclassification by an IDS, but also satisfy domain constraints. Instead of perturbing the data samples itself, the adversarial examples are crafted by perturbing the latent space representation of the data sample. It allows the generation of adversarial examples under limited perturbation. This research has explored the novel applications of generative networks for generating constraint satisfying adversarial examples. The experimental results support the claims with an attack success rate of 64.8\(\%\) against ML/DL-based IDS. The trained model can be integrated further into an operational IDS to strengthen its robustness against adversarial examples; however, this is out of scope of this work.