Global Sensitivity
What you'll learn: You'll understand how global sensitivity quantifies the maximum impact a single data record can have on a query's output, which directly determines how much noise differential privacy mechanisms must add.
The Core Idea
Global sensitivity measures the worst-case change in a query or function's output when you add, remove, or modify exactly one record in your dataset. Think of it as asking: "What's the biggest difference one person's data could possibly make?"
This concept is fundamental to differential privacy because it tells us how much noise we need to add to mask the influence of any individual. If one record can swing a result by a large amount, we need proportionally more noise to hide its contribution.
Why It Matters
The Laplace mechanism (which you've just learned) adds noise calibrated to sensitivity. Specifically, noise is drawn from a distribution with scale Δf/ε, where:
Δfis the global sensitivityεis your privacy budget
Lower sensitivity means less noise needed for the same privacy guarantee—your privatized results will be more accurate.
A Simple Example
Consider counting how many people in a dataset have a certain attribute. Adding or removing one person changes the count by at most 1. So the global sensitivity is 1.
For computing an average age, if ages range from 0 to 120 and your dataset has n people, one person's contribution is at most 120/n. But sensitivity also accounts for changing the denominator—making calculations more complex.
Key Takeaway: Global sensitivity is the maximum possible change to a query output from any single record modification, and it determines the noise scale required to achieve differential privacy.