Research

Paper

AI LLM February 23, 2026

VALD: Multi-Stage Vision Attack Detection for Efficient LVLM Defense

Authors

Nadav Kadvil, Ayellet Tal

Abstract

Large Vision-Language Models (LVLMs) can be vulnerable to adversarial images that subtly bias their outputs toward plausible yet incorrect responses. We introduce a general, efficient, and training-free defense that combines image transformations with agentic data consolidation to recover correct model behavior. A key component of our approach is a two-stage detection mechanism that quickly filters out the majority of clean inputs. We first assess image consistency under content-preserving transformations at negligible computational cost. For more challenging cases, we examine discrepancies in a text-embedding space. Only when necessary do we invoke a powerful LLM to resolve attack-induced divergences. A key idea is to consolidate multiple responses, leveraging both their similarities and their differences. We show that our method achieves state-of-the-art accuracy while maintaining notable efficiency: most clean images skip costly processing, and even in the presence of numerous adversarial examples, the overhead remains minimal.

Metadata

arXiv ID: 2602.19570
Provider: ARXIV
Primary Category: cs.CV
Published: 2026-02-23
Fetched: 2026-02-24 04:38

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2602.19570v1</id>\n    <title>VALD: Multi-Stage Vision Attack Detection for Efficient LVLM Defense</title>\n    <updated>2026-02-23T07:39:43Z</updated>\n    <link href='https://arxiv.org/abs/2602.19570v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2602.19570v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Large Vision-Language Models (LVLMs) can be vulnerable to adversarial images that subtly bias their outputs toward plausible yet incorrect responses. We introduce a general, efficient, and training-free defense that combines image transformations with agentic data consolidation to recover correct model behavior. A key component of our approach is a two-stage detection mechanism that quickly filters out the majority of clean inputs. We first assess image consistency under content-preserving transformations at negligible computational cost. For more challenging cases, we examine discrepancies in a text-embedding space. Only when necessary do we invoke a powerful LLM to resolve attack-induced divergences. A key idea is to consolidate multiple responses, leveraging both their similarities and their differences. We show that our method achieves state-of-the-art accuracy while maintaining notable efficiency: most clean images skip costly processing, and even in the presence of numerous adversarial examples, the overhead remains minimal.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CV'/>\n    <published>2026-02-23T07:39:43Z</published>\n    <arxiv:primary_category term='cs.CV'/>\n    <author>\n      <name>Nadav Kadvil</name>\n    </author>\n    <author>\n      <name>Ayellet Tal</name>\n    </author>\n  </entry>"
}