Course contentsShow
← Security
Lesson 2916 of 3,05254. Privacy EngineeringFree lesson

The Gaussian Mechanism and Advanced Noise

Using Gaussian noise for differential privacy, understanding when to use Laplace vs Gaussian, and approximate DP.

The Gaussian Mechanism and Advanced Noise

What you'll learn: How Gaussian noise provides differential privacy with different tradeoffs than Laplace noise, and when to use each mechanism.

The Gaussian Alternative

While the Laplace mechanism adds noise proportional to ε (your privacy budget), the Gaussian mechanism uses noise from a normal distribution. It provides what's called approximate differential privacy (or (ε,δ)-DP), where δ represents a small probability that the stronger privacy guarantee might fail.

Think of it like this: Laplace noise is a strict bouncer—nobody gets through without proper clearance. Gaussian noise is slightly more relaxed, allowing a tiny probability (δ) of information leakage in exchange for better utility.

The Math Behind Gaussian Noise

The Gaussian mechanism adds noise drawn from N(0, σ²) where the variance σ² depends on your sensitivity Δf, privacy parameters ε and δ:

σ ≥ Δf · √(2 ln(1.25/δ)) / ε

This provides (ε,δ)-differential privacy rather than pure ε-DP.

Laplace vs. Gaussian: When to Choose

Use Laplace when:

  • You need pure differential privacy (δ = 0)
  • Working with smaller datasets
  • Regulatory requirements demand strict guarantees

Use Gaussian when:

  • You can tolerate approximate DP (δ > 0, typically 10⁻⁵ to 10⁻¹⁰)
  • You need better utility/accuracy
  • Working with machine learning (DP-SGD uses Gaussian noise)
  • Composition bounds are tighter with advanced composition theorems

Approximate DP in Practice

The δ parameter is kept cryptographically small—often smaller than the probability of hardware failure. For a dataset with n records, common practice sets δ ≈ 1/n² to ensure negligible risk.

Key Takeaway: Gaussian noise offers better utility than Laplace for approximate differential privacy, making it the preferred choice for machine learning applications where small δ is acceptable.