PrivMark: Pixel-Space Steering against Privacy Leakage in Vision-Language Models

Summary

PrivMark is a framework that suppresses soft-biometric leakage, such as age, gender, and race, from vision-language models by adding a bounded pixel-space perturbation to the image. It combines two steering vector directions in hidden-state space, one for incidental disclosure under open-ended descriptions and one for compliance with direct biometric probes. Across three white-box source models it reaches up to 85.9% biometric suppression while keeping 75% of clean OK-VQA utility, and the perturbations transfer to seven unseen VLMs, including closed-source APIs.

Topics

• Multimodal models

Related items

• MediumCVE-2026-121938: langchain4j-aideepin before 3.31.0 contains an insecure direct object reference vulnerability that allows authenticated…Similar attack · NVD/CVE Database
• InfoResidual Attribute Leakage Is an Interface: Auditing Privacy Evaluation After Text De-identificationSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• MediumHundreds of Mundane AI Apps Are Hiding Sexually Abusive Deepfake FeaturesSimilar attack · Wired (Security)
• MediumNew VOLTA Infostealer Uses Claude, Gemini, and Kimi AI to Rank Victims by Financial ValueSimilar attack · The Hacker News
• InfoAccessed, never disclosed: Silent access to protected health information by medical artificial intelligence agentsSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)