Towards Large Language Model (LLM) Forensics Using LLM-based Invocation Log Analysis

Towards Large Language Model (LLM) Forensics Using LLM-based Invocation Log Analysis

연구 분야: Safety

학회: LAMPS '24: Proceedings of the 1st ACM Workshop on Large AI Systems and Models with Privacy and Safety Analysis

Large Language Models (LLMs) have fostered the emergence of software application architectures that improve user experiences powered by generative artificial intelligence. A range of cyber attacks are possible against an LLM. A novel approach to digital forensic analysis of LLM-integrated applications is presented for prompt injection attacks. The forensic analysis process is invoked through LLM log analysis. We propose LLM invocation logging as a critical component for enhancing digital forensic readiness in LLM-integrated applications and evaluate 13 state-of-the-art LLMs for this analysis task. Our findings demonstrate the potential utility of selected LLMs in the context of prompt-to-SQL attacks, influenced by sampling temperature and context window size parameters. We also identify limitations of our work and propose key areas for future research, for ongoing contribution to the emerging field of LLM forensics.