Course contentsShow
← Machine Learning and Deep Learning
Lesson 3381 of 3,53874. Adversarial RobustnessFree lesson

Transferability of Adversarial Examples

How adversarial examples crafted for one model often fool other models, even across architectures.

Transferability of Adversarial Examples

What you'll learn: How adversarial examples crafted to fool one model often successfully fool completely different models, even with different architectures.

The Surprising Cross-Model Vulnerability

You've learned that adversarial examples are carefully perturbed inputs that cause misclassification. Here's something remarkable: an adversarial example created specifically to fool Model A will often also fool Model B, even if B has a completely different architecture, was trained on different data, or uses different activation functions.

This phenomenon is called transferability, and it's both fascinating and concerning for security.

Why Does This Happen?

Transferability suggests that different models learn similar decision boundaries in certain regions of input space. Think of it like multiple architects independently designing buildings—they might use different styles and materials, but they're all constrained by the same laws of physics and human ergonomics. Similarly, neural networks trained on the same task often develop analogous vulnerabilities.

The perturbations exploit fundamental properties of how neural networks partition the input space, not just quirks of one specific model. This is especially true when models are trained on similar datasets or learn similar features.

Security Implications

Transferability makes black-box attacks practical. An attacker doesn't need access to your model's internals—they can:

  1. Train or obtain a substitute model
  2. Generate adversarial examples for it
  3. Use those same examples against your production model

This dramatically lowers the barrier for adversarial attacks in real-world systems where direct model access isn't available.

Key Takeaway: Adversarial examples transfer across models because different networks learn similar vulnerable decision boundaries, enabling practical black-box attacks without direct model access.