2603.20777

Total: 1

#1 OmniPatch: A Universal Adversarial Patch for ViT-CNN Cross-Architecture Transfer in Semantic Segmentation [PDF1] [Copy] [Kimi1] [REL]

Authors: Aarush Aggarwal, Akshat Tomar, Amritanshu Tiwari, Sargam Goyal

Robust semantic segmentation is crucial for safe autonomous driving, yet deployed models remain vulnerable to black-box adversarial attacks when target weights are unknown. Most existing approaches either craft image-wide perturbations or optimize patches for a single architecture, which limits their practicality and transferability. We introduce OmniPatch, a training framework for learning a universal adversarial patch that generalizes across images and both ViT and CNN architectures without requiring access to target model parameters.

Subjects: Machine Learning , Artificial Intelligence , Computer Vision and Pattern Recognition

Publish: 2026-03-21 12:07:42 UTC