Course contentsShow
← Machine Learning and Deep Learning
Lesson 3398 of 3,53874. Adversarial RobustnessFree lesson

Physical-World Adversarial Examples

Study constraints and techniques for creating adversarial examples that remain effective after printing or environmental transformation.

Physical-World Adversarial Examples

What you'll learn: How adversarial examples can survive the transition from digital to physical environments and remain effective under real-world transformations.

The Bridge from Digital to Physical

Up until now, we've explored adversarial examples that exist purely in digital form—carefully crafted pixel perturbations designed to fool models. But what happens when you print that adversarially modified stop sign and mount it on a street corner? Physical-world adversarial examples are designed to remain effective after undergoing transformations like printing, photography, lighting changes, viewing angles, and environmental conditions.

Why the Physical World Is Harder

Think of it like designing a message that must survive being photocopied, crumpled, and photographed from different angles—and still deliver its intended meaning. Physical adversarial examples face similar challenges:

  • Printing and capture: Digital perturbations must survive the printer's color gamut limitations and camera sensor noise
  • Environmental transformations: Changes in lighting, shadows, weather conditions
  • Geometric transformations: Viewing angles, distance, rotation, occlusion
  • Material properties: Texture, reflectance, and surface characteristics

Constraints and Techniques

Creating robust physical adversarial examples requires:

  1. Expectation over Transformations (EOT): During optimization, simulate multiple transformations (rotations, lighting changes, distances) and ensure the perturbation works across all of them
  2. Print-capture simulation: Model the entire print-and-photograph pipeline during adversarial generation
  3. Smoothness constraints: Ensure perturbations don't rely on single-pixel precision that printers can't reproduce
  4. Physically realizable perturbations: Constrain modifications to printable colors and patterns

Key Takeaway: Physical-world adversarial examples require optimization under realistic transformation distributions, ensuring perturbations remain effective across printing, environmental conditions, and viewing variations—making them far more constrained but significantly more dangerous in real deployments.