Dynamic Token Reweighting for Robust Vision-Language Models

TL;DR

This paper introduces DTR, an inference-time method that dynamically reweights visual tokens in vision-language models to defend against multimodal jailbreak attacks without needing additional safety data.

Contribution

DTR is a novel approach that optimizes KV caches at inference time to improve robustness against adversarial visual inputs in VLMs.

Findings

DTR outperforms existing defenses in attack robustness.

DTR maintains high performance on benign tasks.

DTR effectively mitigates multimodal jailbreak attacks.

Abstract

Large vision-language models (VLMs) are highly vulnerable to multimodal jailbreak attacks that exploit visual-textual interactions to bypass safety guardrails. In this paper, we present DTR, a novel inference-time defense that mitigates multimodal jailbreak attacks through optimizing the model's key-value (KV) caches. Rather than relying on curated safety-specific data or costly image-to-text conversion, we introduce a new formulation of the safety-relevant distributional shift induced by the visual modality. This formulation enables DTR to dynamically adjust visual token weights, minimizing the impact of adversarial visual inputs while preserving the model's general capabilities and inference efficiency. Extensive evaluation across diverse VLMs and attack benchmarks demonstrates that DTR outperforms existing defenses in both attack robustness and benign-task performance, marking the first successful application of KV cache optimization for safety enhancement in multimodal foundation models. The code for replicating DTR is available at: https://github.com/TanqiuJiang/DTR.