AI Paper Atlas

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

VCP-Attack: Visual-Contrastive Projection for Transferable Black-Box Targeted Attacks on Large Vision-Language Models

Jiawei Zhao, Minjie Du, Zihan Qin, Zhuoran Wang, Lizhe Xie, Yining Hu

CVPR 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: Vision-language model robustness

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
add PCA subspace projection (d = 10) (ablation)VCP-Attack · Open-source LVLMs avgASR (%): 90.5 -> 96.7Improved
use PCA dimension 10 instead of 5, 30 or 50 (ablation)VCP-Attack · Open-source LVLMs avgASR (%)Improved
set positive loss coefficient alpha = 0.8 mixing top-k and max terms (ablation)VCP-Attack · Open-source LVLMs avgASR (%)Improved
use top-k/top-m = (10,10) sampling instead of larger or no filtering (ablation)VCP-Attack · Open-source LVLMs avgASR (%)Improved
increase perturbation budget from 8/255 to 32/255 (ablation)VCP-Attack · Open-source LVLMs avgASR (%): up to 98.2Improved
VCP-Attack vs strongest baseline (M-Attack)VCP-Attack · Open-source LVLMs avgASR (%): 71.8 -> 94.2Improved
VCP-Attack vs strongest baseline on Claude-3.7-SonnetVCP-Attack · Claude-3.7-SonnetASR (%): 38.8 -> 61.1Improved
scale victim model size from 14B to 38B/78BM-Attack and FOA-Attack · InternVL3ASR (%): 87.7 -> 54.4 (M-Attack)Hurt

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