Reproduce “Vulnerability Analysis of Transformer-based Optical Character Recognition to Adversarial Attacks” 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
Lucas Beerens, Desmond J. Higham
The research by Lucas Beerens and Desmond J. Higham investigates the susceptibility of Transformer-based Optical Character Recognition (TrOCR) models to adversarial attacks. It demonstrates that TrOCR is vulnerable to visually imperceptible image perturbations, which can lead to significant character error rates in untargeted attacks and successful targeted misclassifications.
Key insights
- TrOCR models, despite their advanced architecture, exhibit inherent fragility to adversarial perturbations, similar to other deep neural networks.
- The effectiveness of adversarial attacks on TrOCR varies by method; C&W proved more potent for generating small, imperceptible untargeted and targeted perturbations, while FGSM caused more drastic text corruption with larger perturbations.
- The identified vulnerabilities highlight a potential compliance gap with emerging AI regulations, such as the EU AI Act, which mandates resilience for high-risk AI systems against unauthorized alterations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.