Reproduce “TopoReformer: Mitigating Adversarial Attacks Using Topological Purification in OCR 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
Bhagyesh Kumar, A S Aravinthakashan, Akshat Satyanarayan, Ishaan Gakhar, Ujjwal Verma
Researchers from Manipal Institute of Technology developed TopoReformer, a defense pipeline for OCR models that uses topological data analysis to purify adversarial examples. It effectively removes subtle perturbations while preserving global structural integrity, achieving improved robustness against various attacks while maintaining high performance on unperturbed inputs.
Key insights
- Persistent homology can effectively identify and preserve the global structural integrity of data, providing a robust mechanism to purify adversarial examples distinct from pixel-level denoising.
- A multi-stage defense combining topology-aware feature extraction with a manifold-aligning VAE and a specialized latent injection mechanism improves robustness against a wide range of attacks.
- Enforcing topological consistency during purification leads to a more stable and semantically consistent latent manifold, which can enhance classifier confidence on both clean and adversarial inputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.