Reproduce “GeoDetect: Geometric Adversarial Detection for VLPs” 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
Afsaneh Hasanebrahimi, Hanxun Huang, Christopher Leckie, James Bailey, Sarah Erfani
GeoDetect, a new method for detecting adversarial examples in Vision-Language Pre-trained models (VLPs), leverages geometric distortions caused by adversarial perturbations in VLP embedding spaces. It achieves high detection performance, with AUCs up to 100% and FPR95 of 0% on some datasets, across various VLP architectures and multimodal tasks like zero-shot classification and image-text retrieval.
Key insights
- VLP embedding spaces exhibit strong anisotropy, a property that adversarial perturbations significantly alter, causing embeddings to deviate into sparser, higher-dimensional regions.
- Adversarial examples consistently demonstrate greater expected distances to randomly sampled clean embeddings compared to distances between clean embeddings themselves.
- Adversarial perturbations suppress tangent components and amplify normal components, effectively moving VLP embeddings off the clean data manifold.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.