GCG Attack On A Diffusion LLM

TL;DR

This paper explores the vulnerability of diffusion-based large language models to GCG-style adversarial prompt attacks, providing initial insights into their robustness and highlighting the need for new defense strategies.

Contribution

It is the first to analyze GCG attack applicability to diffusion LLMs, evaluating various attack variants on LLaDA and assessing their robustness.

Findings

GCG-style attacks can successfully generate adversarial prompts for diffusion LLMs.

Diffusion LLMs show varying robustness depending on attack type and prompt perturbation.

The study highlights the need for developing new defense mechanisms against adversarial prompts.

Abstract

While most LLMs are autoregressive, diffusion-based LLMs have recently emerged as an alternative method for generation. Greedy Coordinate Gradient (GCG) attacks have proven effective against autoregressive models, but their applicability to diffusion language models remains largely unexplored. In this work, we present an exploratory study of GCG-style adversarial prompt attacks on LLaDA (Large Language Diffusion with mAsking), an open-source diffusion LLM. We evaluate multiple attack variants, including prefix perturbations and suffix-based adversarial generation, on harmful prompts drawn from the AdvBench dataset. Our study provides initial insights into the robustness and attack surface of diffusion language models and motivates the development of alternative optimization and evaluation strategies for adversarial analysis in this setting.