Reproduce “Towards Robustness against Typographic Attack with Training-free Concept Localization” 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
Bohan Liu, Wenqian Ye, Guangzhi Xiong, Zhenghao He, Sanchit Sinha, Aidong Zhang
Researchers at the University of Virginia introduced a training-free mechanistic interpretability method that pinpoints and mitigates typographic attack vulnerabilities within the attention heads of CLIP-based models. The approach improved object classification accuracy under attack by an average of 71.7% and enhanced Visual Question Answering accuracy on attacked images by up to 2.11%, while maintaining general performance.
Key insights
- Typographic attacks are rooted in specific, identifiable "typographic reading circuits" within the Multi-Head Self-Attention (MHSA) modules of Vision Transformers.
- Stochastic sampling in the lower-dimensional MHSA head subspaces can efficiently and effectively localize lexical-focused concepts without extensive training or dictionary learning.
- Precise, training-free interventions applied to these localized circuits offer a potent strategy for enhancing robustness against typographic attacks with minimal impact on general model performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.