Course contentsShow
← Security
Lesson 2814 of 3,05253. AI and ML SecurityFree lesson

Physical World Adversarial Examples

Creating robust perturbations that survive camera capture, printing, and environmental variations in real deployments.

Physical World Adversarial Examples

What you'll learn: How adversarial perturbations can be engineered to survive real-world transformations like printing, camera capture, and environmental variations—turning digital attacks into physical threats.

From Digital to Physical

In previous lessons, you learned to craft adversarial examples that fool ML models in the digital domain. But what happens when you print that adversarial image, photograph it with a phone camera, or place it in varying lighting conditions? Most digital perturbations fail immediately. Physical world adversarial examples solve this by creating robust perturbations that remain effective through the entire physical pipeline.

The Challenge of the Real World

Think of digital adversarial examples like whispering a secret password—it works in a quiet room but fails in a noisy crowd. Physical deployment introduces:

  • Print-capture transformations: Color shifts, resolution loss, texture changes
  • Camera artifacts: Lens distortion, auto-focus, compression
  • Environmental variations: Lighting angles, shadows, reflections, distance
  • Viewing angles: The perturbation must work from multiple perspectives

Creating physical adversarials means optimizing perturbations that survive all these transformations simultaneously.

Expectation Over Transformations (EOT)

The key technique is Expectation Over Transformations: instead of optimizing against a single image, you optimize against the expected outcome across many simulated physical transformations (brightness changes, rotations, blur, noise). The perturbation becomes robust because it's been "battle-tested" against real-world variability during creation.

This is why adversarial stop signs, glasses that fool facial recognition, and adversarial patches printed on clothing can actually work in practice—they're designed with physical robustness baked in from the start.

Key Takeaway: Physical world adversarial examples require optimizing perturbations to survive print-capture pipelines and environmental variations, typically using Expectation Over Transformations to ensure robustness across real-world conditions—making ML attacks deployable beyond the digital realm.