A new research paper titled "Curvature Under Attack in hZACH-ViT" explores the interaction of curvature with various factors in a compact Vision Transformer model. The study, conducted on MedMNIST datasets, found that while Poincare geometry generally showed lower adversarial attack success rates, it also resulted in lower clean accuracy. Reducing Poincare curvature improved clean accuracy but significantly increased adversarial attack success on specific datasets. The research suggests that curvature, scale, and proximity to the Poincare boundary collectively influence both clean recognition and adversarial robustness in these models. AI
IMPACT Investigates factors influencing adversarial robustness in vision transformers, potentially informing future model development for improved security.
RANK_REASON Research paper published on arXiv detailing findings about a specific model's adversarial robustness. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Athanasios Angelakis
- CE+DLR
- DagsHub
- Hugging Face
- hZACH-ViT
- MedMNIST
- Projected Gradient Descent
- vision transformer
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →