Course contentsShow
← Machine Learning and Deep Learning
Lesson 3395 of 3,53874. Adversarial RobustnessFree lesson

Black-Box Attacks: Transfer-Based

Leverage transferability of adversarial examples across models to attack systems without gradient access.

Black-Box Attacks: Transfer-Based

You'll learn how attackers exploit adversarial example transferability to compromise models without needing internal access.

The Core Idea

In a black-box setting, you can't access a model's architecture, weights, or gradients—you only see predictions. Transfer-based attacks solve this problem by leveraging a key property you've already learned: adversarial examples transfer across different models.

Here's how it works: an attacker trains or obtains a substitute model (a proxy they control completely), generates adversarial examples against that substitute using white-box methods like FGSM or PGD, then feeds those same perturbed inputs to the target model they want to attack. Surprisingly often, the adversarial examples fool both models.

Why This Works

Models trained on similar data often learn comparable decision boundaries, even with different architectures. When you craft an adversarial example that crosses the decision boundary of your substitute model, there's a good chance it also crosses the boundary of the target model—especially if they share similar training data or task domains.

Think of it like this: If you learn to fool one security guard by wearing a specific disguise, that same disguise might fool other guards at different buildings who've been trained with similar protocols.

The Attack Process

  1. Query the target to understand its behavior (optional reconnaissance)
  2. Train a substitute model on similar data or using the target's predictions
  3. Generate adversarial examples using white-box attacks on your substitute
  4. Transfer those examples to attack the real target model

This is practical and dangerous because it requires no privileged access to the victim system.

Key Takeaway: Transfer-based black-box attacks exploit adversarial transferability to compromise models using only their input-output behavior, making gradient-free attacks feasible in real-world scenarios.