Course contentsShow
← Machine Learning and Deep Learning
Lesson 3342 of 3,53873. Privacy in Machine LearningFree lesson

The Gaussian Mechanism

Using Gaussian noise for (ε,δ)-differential privacy; relationship between noise variance and privacy parameters.

The Gaussian Mechanism

What you'll learn: How to achieve (ε,δ)-differential privacy using Gaussian noise instead of Laplace noise, and when this approach is appropriate.


While the Laplace Mechanism provides pure ε-differential privacy, many practical scenarios benefit from a slightly relaxed guarantee that allows a small failure probability δ. Enter the Gaussian Mechanism.

Core Idea

The Gaussian Mechanism adds noise drawn from a normal distribution to your query output. Instead of Laplace noise (which has "heavier tails"), you use Gaussian noise with carefully calibrated variance. This trades pure privacy (ε-DP) for approximate privacy ((ε,δ)-DP), but often allows less noise for the same privacy level.

The Formula

For a query with global sensitivity Δf, add Gaussian noise:

noise ~ N(0, σ²)

where the standard deviation σ must satisfy:

σ ≥ Δf · √(2·ln(1.25/δ)) / ε

Key relationships:

  • Higher ε (weaker privacy) → smaller σ → less noise needed
  • Smaller δ (stricter failure bound) → larger σ → more noise required
  • Higher sensitivity Δf → proportionally more noise needed

When to Use It

The Gaussian Mechanism shines when:

  • You can tolerate δ > 0 (common in practice, often set to 1/n² where n is dataset size)
  • You're composing many privacy operations (Gaussian noise composes well)
  • You want tighter utility-privacy tradeoffs than Laplace in high dimensions

Key Takeaway: The Gaussian Mechanism provides (ε,δ)-differential privacy by adding normal noise with variance proportional to sensitivity, privacy budget, and failure probability—offering more flexible privacy-utility tradeoffs than pure differential privacy.