GLEAM: GAN and LLM for Evasive Adversarial Malware

The rapid evolution of cybersecurity threats poses formidable challenges for effective malware detection. Traditional methods often struggle to keep pace with the continuously changing landscape of new malware variants. To address this issue, researchers have turned to machine-learning techniques to optimize malware detection. Yet, these approaches have overlooked the fact that evasive malware is meticulously crafted to elude detection through tactics such as exploiting software vulnerabilities, utilizing encryption, and employing obfuscation techniques. Malware authors have a strong incentive to attack malware detection systems, yet the features and methods that they exploit are limited. We propose a model named GLEAM- GAN and LLM for Evasive Adversarial Malware. This model infuses hex code and opcode features with LLM (Large Language Model) embeddings and GANs (Generative Adversarial Networks) to generate synthetic samples that closely resemble evasive mal-ware to bypass black-box machine learning detectors. Through extensive evaluation, our model achieved an average evasion rate increase of 22.6%, demonstrating its ability to effectively attack detection systems. By expanding the space for adversarial malware generation, we give modern detection systems the capability to counter the nuanced tactics of evasive malware, thus enhancing proficiency in preempting and neutralizing potential threats with heightened precision.