Course contentsShow
← Data Science
Lesson 1900 of 2,14544. Data Ethics and PrivacyFree lesson

Privacy Budget and Composition

Understand epsilon parameters and how privacy degrades with repeated queries using the same data.

Privacy Budget and Composition

You'll learn how repeated queries on the same data consume privacy and why there's a limit to how many analyses you can safely perform.

The Core Idea

In differential privacy, every query that accesses sensitive data "spends" some privacy. The epsilon (ε) parameter represents this cost—lower epsilon means stronger privacy guarantees, but also more noise added to results. The total epsilon across all queries is your privacy budget.

Here's the critical insight: privacy degrades with composition. Each time you run a query on the same dataset, you accumulate risk. Even if each individual query is differentially private, the combined information from multiple queries can potentially reveal more about individuals than intended.

How Composition Works

Think of your privacy budget like a bank account with ε dollars. Each analysis "withdraws" some amount:

  • Simple composition: If you run 10 queries each with ε=0.1, your total privacy loss is ε=1.0 (worst case)
  • Advanced composition: Mathematical techniques can provide tighter bounds, so the total might be less than simple addition
  • Once spent, it's gone: You cannot query indefinitely—eventually you exhaust your budget and must stop

Why This Matters

If a researcher runs 100 slightly different analyses on medical data, they might reconstruct individual records even if each query added noise. Organizations must:

  • Set a total epsilon budget upfront (e.g., ε=1.0 for all analyses)
  • Track epsilon spent per query
  • Reject requests once the budget is exhausted

Key Takeaway: Privacy budget (epsilon) is finite—each query consumes some of it, and composition means repeated queries compound privacy loss, requiring careful tracking and enforcement to protect individuals.