Course contentsShow
← Security
Lesson 2848 of 3,05253. AI and ML SecurityFree lesson

Certified Defenses and Provable Robustness

Techniques that provide mathematical guarantees about model behavior within bounded perturbations.

Certified Defenses and Provable Robustness

What You'll Learn

You'll learn how to mathematically prove that a machine learning model will behave correctly even when adversaries perturb inputs within a specific boundary.

The Core Idea

Adversarial training (the previous lesson) empirically strengthens models by exposing them to adversarial examples during training. But it offers no guarantees—you never know if a cleverer attack might still fool your model.

Certified defenses flip this paradigm. Instead of hoping your model is robust, you mathematically prove that no possible attack within a defined perturbation budget can change the model's prediction. If an input is certified at radius ε, you have a formal guarantee that any perturbation smaller than ε cannot cause misclassification.

Think of it like the difference between testing a bridge with trucks (adversarial training) versus calculating its maximum load capacity from engineering principles (certification). One is empirical, the other is mathematical proof.

How It Works

Common certification techniques include:

  • Randomized smoothing: Add noise to inputs and use statistics to bound prediction stability
  • Interval bound propagation: Track how perturbations propagate through network layers using mathematical intervals
  • Convex relaxation: Approximate neural network behavior with simpler, analyzable functions

These methods compute a certified radius around each input. Within this radius, the prediction is provably unchanged, regardless of how cleverly an attacker chooses their perturbation.

The tradeoff? Certification is computationally expensive, and certified radii are often smaller than empirical robustness from adversarial training. You gain mathematical certainty but sacrifice some practical robustness.

Key Takeaway:
Certified defenses provide mathematical guarantees about model predictions within bounded perturbations, offering provable security rather than empirical robustness—though often at higher computational cost and smaller certified radii.