Reproduce “Private Attribute Inference from Images with Vision-Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Batuhan Tömekçe, Mark Vero, Robin Staab, Martin Vechev
Vision-Language Models (VLMs) can infer sensitive personal attributes from inconspicuous online images with high accuracy, reaching 77.6% with GPT-4V, even when direct human depictions are absent. This capability, coupled with easily circumvented safeguards, enables privacy violations at an unprecedented scale, necessitating urgent attention from model providers, regulators, and users.
Key insights
- VLMs represent a "categorical paradigm shift" in privacy threats, enabling automated and highly scalable inference of sensitive personal attributes from seemingly innocuous visual data with unprecedented efficiency.
- Current safety mechanisms in state-of-the-art VLMs are "incredibly brittle" and easily circumvented by even basic prompt engineering, demonstrating a significant vulnerability.
- The ability to infer private attributes strongly correlates with the general capabilities of VLMs, indicating that future, more powerful models will pose an escalating and more pervasive privacy risk.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.