Large Language Models in Side-Channel Cryptanalysis

Authors

  • Witold Waligóra Myre Laboratories

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 semi-autonomously 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.

Additional Files

Published

2025-05-30

Issue

Section

Cryptography and Cybersecurity