Reproduce “Semantic Adversarial Examples” 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
Hossein Hosseini, Radha Poovendran
Researchers at the University of Washington introduce "Semantic Adversarial Examples" by uniformly shifting colors in the HSV space, demonstrating that state-of-the-art deep neural networks experience severe accuracy drops (e.g., VGG16 from 93.6% to 5.7%) while the images remain semantically identical and natural-looking to human observers. This work reveals a fundamental fragility in how models perceive visual information compared to humans and bypasses existing adversarial defenses.
Key insights
- Deep neural networks, including those designed for adversarial robustness, are highly vulnerable to semantic transformations like color shifts, indicating a reliance on superficial color features rather than robust semantic understanding.
- Existing adversarial defense mechanisms, primarily focused on imperceptible perturbations, are largely ineffective against these new semantic adversarial examples.
- A significant cognitive gap exists between human and machine perception, where humans are invariant to such color shifts for object recognition, but CNNs are fundamentally sensitive.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.