AI Paper Atlas

Research map / Adversarial robustness and security / Attacks on text and graphs

Localize and Neutralize: Gradient-Guided Token Suppression Against Visual Prompt Injection Attack

Dongpeng Zhang, Ke Ma, Yangbangyan Jiang, Gaozheng Pei, Longtao Huang, Qianqian Xu, Qingming Huang

ICML 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: Vision-language model robustness · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use hidden-state gradient norm instead of output logit gradient for token attribution (ablation)GTM · jailbreak, aligned first tokenattack success rate (%): 30.91 -> 24.55Improved
use hidden-state gradient norm instead of logit gradient (ablation)GTM · misaligned first tokenattack success rate (%): 0.91 -> 0.91No clear effect
raise masked token ratio beyond 5% (ablation)GTM · ASR and MM-VetASR and utilityMixed
apply GTM defenseQwen2-VL-7B · VMA attacksaverage ASR (%): 90.54 -> 0.00Improved
apply GTM defenseLLaVA-v1.5-7b · VMA attacksaverage ASR (%): 99.64 -> 1.27Improved
use GTM instead of DiffPuredefense · MM-Vet utility and costutility, time, memory: 7.22 s -> 3.50 sImproved
use GTM instead of Cider under unbounded perturbationjailbreak defense · jailbreak eps=infattack success rate (%): 94.55 -> 0.91Improved
apply GTM against optimization-based attacks APGDGTM · APGD eps=32attack success rate: 6 vs Guard 6No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy