Attacking Large Language Models with Projected Gradient Descent

TL;DR

This paper introduces an improved PGD-based method for attacking large language models efficiently, significantly reducing computational costs while maintaining high attack success rates.

Contribution

It demonstrates that controlling relaxation errors enhances PGD attack effectiveness, making it a faster alternative to discrete optimization for adversarial prompts.

Findings

PGD attacks become more effective with controlled relaxation errors

The proposed method is up to ten times faster than discrete optimization

Achieves comparable or better attack success rates

Abstract

Current LLM alignment methods are readily broken through specifically crafted adversarial prompts. While crafting adversarial prompts using discrete optimization is highly effective, such attacks typically use more than 100,000 LLM calls. This high computational cost makes them unsuitable for, e.g., quantitative analyses and adversarial training. To remedy this, we revisit Projected Gradient Descent (PGD) on the continuously relaxed input prompt. Although previous attempts with ordinary gradient-based attacks largely failed, we show that carefully controlling the error introduced by the continuous relaxation tremendously boosts their efficacy. Our PGD for LLMs is up to one order of magnitude faster than state-of-the-art discrete optimization to achieve the same devastating attack results.