Harnessing LLMs for IoT Malware Detection: A Comparative Analysis of BERT and GPT-2

In recent years, the proliferation of Internet of Things (IoT) devices has introduced significant vulnerabilities in cybersecurity, particularly with the rise of sophisticated malware targeting these systems. Traditional detection methods, often based on static signatures, struggle to keep pace with evolving threats, such as zero-day attacks. This paper explores the application of Large Language Models (LLMs), specifically BERT and GPT-2, in detecting IoT malware by analyzing network traffic and identifying anomalies. Using the contextual understanding and adaptability of LLM, our approach significantly enhances detection accuracy compared to conventional methods. We evaluated the models using the ToN-IoT dataset, demonstrating their capability to detect complex malware patterns with higher precision. The results indicate that BERT outperforms GPT-2 across multiple metrics, highlighting its effectiveness in generalizing to various attack types. Despite promising advancements, challenges such as computational resource demands and model interpretability persist. Future research should focus on optimizing LLMs for real-time detection in resource-constrained environments and improving transparency to enhance trust among cybersecurity professionals. Our study underscores the potential of LLMs as powerful tools in the ongoing battle against IoT malware, offering a robust framework for enhancing cybersecurity defenses.