OmniPatch: A Universal Adversarial Patch for ViT-CNN Cross-Architecture Transfer in Semantic Segmentation

Universal adversarial patches pose a significant threat to autonomous driving systems: OmniPatch demonstrates that a single adversarial patch can fool both Vision Transformer (ViT) and CNN-based semantic segmentation models without needing access to the target model's internal parameters, raising serious safety concerns for deployed autonomous vehicles.

Black-box attacks are a critical vulnerability: The research shows that adversarial attacks work effectively even when attackers have no knowledge of the target model's weights or architecture, meaning real-world deployed systems cannot rely on security through obscurity.

Cross-architecture transferability makes defenses more challenging: Because the same patch transfers across different model types (ViTs and CNNs), defenders cannot simply switch between architectures as a mitigation strategy—a common assumption in AI safety that this work invalidates.

Semantic segmentation systems require specific robustness improvements: Unlike image classification, semantic segmentation is critical for autonomous driving because it must correctly identify road elements across entire scenes; the paper highlights that this task remains particularly vulnerable to adversarial manipulation.

Practical adversarial threats require urgent attention: By demonstrating that a single physical patch can attack multiple model types without target knowledge, this work emphasizes that adversarial robustness must be a priority during model development and deployment, not an afterthought.