Empowering Hardware Security with LLM: The Development of a Vulnerable Hardware Database

The scarcity of comprehensive databases and bench-marks in hardware design specifically tailored for security tasks is a significant challenge in the community. Such databases are crucial for developing machine learning-based methods and benchmarking, providing a foundation for evaluating and improving hardware security solutions. However, manually creating these extensive datasets is impractical due to the significant time and effort required. Given the proficiency of large language models (LLM) in natural language processing, coding, and advanced reasoning tasks, using LLM as an artificial intelligence (AI) agent presents a viable option to efficiently create such extensive datasets. In this light, this paper introduces Vul-FSM, a database of 10,000 vulnerable finite state machine (FSM) designs incorporating 16 distinct security weaknesses and vulnerabilities generated using the proposed SecRT-Llmframework. The framework combines the in-context learning capability of LLM, the guidance of developed prompting strategies, and the scrutiny of fidelity-check to not only insert but also detect hardware vulnerabilities and weaknesses. To demonstrate the efficacy of SecRT-LLM, we present an exhaustive analysis, highlighting the proficiency of GPT models in vulnerability insertion, detection, and mitigation. Our proposed SecRT-LLM framework, using gpt-3.5-turbo, demonstrates strong effectiveness, achieving macroaverage pass rates of 81.98% and 80.30% on the first attempt and 97.37% and 99.07% within five attempts for vulnerability insertion and detection, respectively.