Physical-World Adversarial Examples
What you'll learn: How adversarial perturbations can survive the leap from digital screens into the real world through printing, photography, and environmental conditions.
From Pixels to Pavement
You've learned that adversarial examples fool models with tiny digital perturbations. But what happens when you print that adversarially-crafted image and photograph it? Surprisingly, many adversarial perturbations survive this physical transformation—and that's a serious security concern.
Physical-world adversarial examples are perturbations designed to remain effective even after:
- Printing (ink/toner artifacts, color shifts)
- Photography (camera angle, lighting, distance)
- Environmental variations (weather, shadows, reflections)
Why This Matters
Think of a stop sign with carefully crafted stickers. In the digital realm, you might tweak individual pixels imperceptibly. In the physical world, you need perturbations robust enough to survive real-world transformations. Researchers have demonstrated adversarial patches on stop signs that cause autonomous vehicles to misclassify them as speed limit signs—even when photographed from different angles and distances.
The key challenge is robustness under transformation. Unlike digital adversarial examples that exist in perfect pixel space, physical examples must account for:
- Variable viewing angles
- Changing illumination
- Camera sensor noise
- Lens distortion
- Partial occlusion
Expectation Over Transformation
To create physical adversarial examples, attackers often optimize perturbations over an ensemble of transformations—simulating different angles, lighting conditions, and distances during the crafting process. This makes the perturbation effective across the range of real-world conditions it might encounter.
Key Takeaway: Physical-world adversarial examples demonstrate that adversarial vulnerabilities aren't just theoretical—they can manifest in real environments, surviving printing and photography to threaten deployed systems like autonomous vehicles and facial recognition gates.