Adversarial Tuning: Defending Against Jailbreak Attacks for LLMs

TL;DR

This paper introduces a two-stage adversarial tuning framework to improve the robustness of Large Language Models against jailbreak attacks by generating worst-case adversarial prompts and refining defenses.

Contribution

It presents a novel hierarchical meta-universal adversarial prompt learning method and an automatic prompt refinement process to enhance LLMs' defense against jailbreak attacks.

Findings

Outperforms six baseline defenses across three datasets

Demonstrates robustness against multiple attack scenarios

Shows transferability across different LLMs

Abstract

Although safely enhanced Large Language Models (LLMs) have achieved remarkable success in tackling various complex tasks in a zero-shot manner, they remain susceptible to jailbreak attacks, particularly the unknown jailbreak attack. To enhance LLMs' generalized defense capabilities, we propose a two-stage adversarial tuning framework, which generates adversarial prompts to explore worst-case scenarios by optimizing datasets containing pairs of adversarial prompts and their safe responses. In the first stage, we introduce the hierarchical meta-universal adversarial prompt learning to efficiently and effectively generate token-level adversarial prompts. In the second stage, we propose the automatic adversarial prompt learning to iteratively refine semantic-level adversarial prompts, further enhancing LLM's defense capabilities. We conducted comprehensive experiments on three widely used jailbreak datasets, comparing our framework with six defense baselines under five representative attack scenarios. The results underscore the superiority of our proposed methods. Furthermore, our adversarial tuning framework exhibits empirical generalizability across various attack strategies and target LLMs, highlighting its potential as a transferable defense mechanism.