Course contentsShow
← Machine Learning and Deep Learning
Lesson 3385 of 3,53874. Adversarial RobustnessFree lesson

Adversarial Patches

Localized perturbations that can be placed anywhere in an image to cause misclassification.

Adversarial Patches

What you'll learn: How adversarial perturbations can be concentrated into small, localized patches that fool models anywhere in an image.

The Core Idea

Unlike traditional adversarial examples that modify an entire image with imperceptible noise, adversarial patches are small, visible regions that can be placed anywhere in an image to cause misclassification. Think of it like a sticker: you can print it out, stick it on an object, and the model will suddenly misclassify that object—regardless of where you place the sticker.

The key insight is that these patches don't need to cover the whole image. A strategically designed patch occupying just 5-10% of the image can completely fool a classifier, even when the rest of the image is perfectly normal and unmodified.

Why This Matters

Adversarial patches are particularly dangerous in the physical world because:

  • Easy to deploy: Print the patch, stick it anywhere
  • Location-independent: Work regardless of where they appear
  • Visible but effective: Even though humans can see them, models still fail
  • Universal: A single patch can fool the model on many different images

Real-World Example

Imagine a printed sticker designed to make any stop sign get classified as a speed limit sign. An attacker could place this sticker on actual stop signs, and autonomous vehicles using vulnerable models might dangerously misinterpret the sign.

The optimization process creates patches that maximize classification error across many images and positions, making them robust to viewing angles, lighting, and placement.

Key Takeaway: Adversarial patches demonstrate that adversarial robustness failures can be concentrated into small, physical objects that work universally across scenes, making them a practical security threat beyond digital attacks.