UL-VAE: An Unsupervised Learning Approach for Zero-day Malware Detection Using Variational Autoencoder

UL-VAE: An Unsupervised Learning Approach for Zero-day Malware Detection Using Variational Autoencoder

연구 분야: Strategies

학회: 2024 International Conference on Computational Intelligence and Network Systems (CINS)

With the rapid growth of Internet of Things (IoT) technology, there is an increase in the dangers of malware attacks. The underlying threat in these attacks is invading people's privacy and security through the use of vulnerabilities in IoT devices. The most known malware type is zero-day, which is a newly innovated malware that has no background; hence, no training has been initiated to be undetectable and unstoppable from violent software attacks. Traditional intrusion detection systems (IDS) often face difficulties in identifying these new and evolving threats since they rely on pre-established attack patterns or signatures. This study aims to develop a new method of IoT malware traffic classification using unsupervised learning-based variational autoencoder (UL-VAE). It is demonstrated that UL-VAEs are effective in capturing the latent representations of benign software images. In this manner, zero-day malware can be suspected whenever the represented images deviate from the benign ones. The examination of experimental results is conducted utilising an image-based IoT malware binary dataset. The UL-VAE detects zero-day attacks with an accuracy of 0.9600, a false positive rate (FPR) of 0.0203, and a false negative rate (FNR) of 0.0363, significantly improving some of the existing IDS methodologies. This study emphasises the potential of UL-VAE to enhance cybersecurity strategies against emerging and unknown threats.