Directional Embedding Smoothing for Robust Vision Language Models

TL;DR

This paper introduces a novel embedding smoothing technique called RESTA, which enhances the robustness of vision-language models against jailbreaking attacks by aligning noise with token embeddings, thereby improving security.

Contribution

The paper extends the RESTA defense to VLMs and demonstrates its effectiveness against multi-modal jailbreaking attacks using directional embedding noise.

Findings

RESTA reduces attack success rates on VLMs.

Directional embedding noise improves defense effectiveness.

RESTA is lightweight and suitable for inference-time deployment.

Abstract

The safety and reliability of vision-language models (VLMs) are a crucial part of deploying trustworthy agentic AI systems. However, VLMs remain vulnerable to jailbreaking attacks that undermine their safety alignment to yield harmful outputs. In this work, we extend the Randomized Embedding Smoothing and Token Aggregation (RESTA) defense to VLMs and evaluate its performance against the JailBreakV-28K benchmark of multi-modal jailbreaking attacks. We find that RESTA is effective in reducing attack success rate over this diverse corpus of attacks, in particular, when employing directional embedding noise, where the injected noise is aligned with the original token embedding vectors. Our results demonstrate that RESTA can contribute to securing VLMs within agentic systems, as a lightweight, inference-time defense layer of an overall security framework.