Large Language Models in side-channel cryptanalysis

Abstract

Recent advancements in large language models (LLMs) have demonstrated their potential beyond conventional natural language processing tasks. This study demonstrates that GPT-4, a state-of-the-art large language model, can semiautonomously generate and execute side-channel attacks, specifically Correlation Power Analysis (CPA) and timing attacks. By letting the model build and execute code on physical hardware as well as collect and analyze power traces and timing information I’ll show that a non-expect operator equipped with an LLM can execute CPAs against industry-standard embedded encryption libraries. The findings suggest that LLMs’ capabilities present both opportunities for accelerated research and challenges related to the potential misuse of such technologies.

Keywords

• side-channel analysis
• large language models
• gpt-4