SafeTy Reasoning Elicitation Alignment for Multi-Turn Dialogues

TL;DR

This paper introduces STREAM, a novel safety alignment method that detects malicious multi-turn dialogues to protect large language models from attacks while maintaining their functionality.

Contribution

The paper presents a new safety reasoning moderator trained on a human-annotated dataset to effectively identify malicious intent in multi-turn conversations.

Findings

Reduces attack success rate by 51.2%

Outperforms existing defense techniques

Maintains LLM capabilities

Abstract

Malicious attackers can exploit large language models (LLMs) by engaging them in multi-turn dialogues to achieve harmful objectives, posing significant safety risks to society. To address this challenge, we propose a novel defense mechanism: SafeTy Reasoning Elicitation Alignment for Multi-Turn Dialogues (STREAM). STREAM defends LLMs against multi-turn attacks while preserving their functional capabilities. Our approach involves constructing a human-annotated dataset, the Safety Reasoning Multi-turn Dialogues dataset, which is used to fine-tune a plug-and-play safety reasoning moderator. This model is designed to identify malicious intent hidden within multi-turn conversations and alert the target LLM of potential risks. We evaluate STREAM across multiple LLMs against prevalent multi-turn attack strategies. Experimental results demonstrate that our method significantly outperforms existing defense techniques, reducing the Attack Success Rate (ASR) by 51.2%, all while maintaining comparable LLM capability.