How Jailbreak Defenses Work and Ensemble? A Mechanistic Investigation

TL;DR

This paper investigates how jailbreak defenses work in large vision-language models, identifying key mechanisms and proposing ensemble strategies to improve safety without sacrificing helpfulness.

Contribution

It systematically analyzes jailbreak defenses, introduces two key mechanisms, and develops ensemble strategies to balance safety and helpfulness in LVLMs.

Findings

Ensemble defenses improve safety and helpfulness trade-offs.

Safety shift increases overall refusal rates.

Harmfulness discrimination enhances harmful input detection.

Abstract

Jailbreak attacks, where harmful prompts bypass generative models' built-in safety, raise serious concerns about model vulnerability. While many defense methods have been proposed, the trade-offs between safety and helpfulness, and their application to Large Vision-Language Models (LVLMs), are not well understood. This paper systematically examines jailbreak defenses by reframing the standard generation task as a binary classification problem to assess model refusal tendencies for both harmful and benign queries. We identify two key defense mechanisms: safety shift, which increases refusal rates across all queries, and harmfulness discrimination, which improves the model's ability to distinguish between harmful and benign inputs. Using these mechanisms, we develop two ensemble defense strategies-inter-mechanism ensembles and intra-mechanism ensembles-to balance safety and helpfulness. Experiments on the MM-SafetyBench and MOSSBench datasets with LLaVA-1.5 models show that these strategies effectively improve model safety or optimize the trade-off between safety and helpfulness.