Epsilon and Privacy Budgets
You'll learn: How the epsilon parameter and privacy budgets mathematically control how much information can leak from individual data points in differentially private systems.
What is Epsilon?
Epsilon (ε) is the privacy loss parameter in differential privacy. It quantifies how much a single person's data can influence the output of a query or model. A smaller epsilon means stronger privacy guarantees—the output looks nearly identical whether or not any individual's data is included.
Think of epsilon like a "privacy knob":
- ε = 0.1 → Very private, minimal information leakage
- ε = 1.0 → Moderate privacy
- ε = 10 → Weak privacy, more accurate results
Lower epsilon adds more noise to outputs, making it harder to reverse-engineer individual contributions but reducing utility.
Privacy Budgets
A privacy budget is the total amount of epsilon you're willing to "spend" across all queries or operations. Each differentially private computation consumes part of this budget.
The catch: Privacy loss accumulates. If you run 10 queries each with ε=0.1, your total privacy loss is approximately ε=1.0. Once you exhaust your budget, you must stop querying or accept weaker privacy guarantees.
Budget Management Strategy
- Set a total budget (e.g., ε=1.0 for the entire dataset lifecycle)
- Allocate fractions to different operations (training: 0.7, validation: 0.2, monitoring: 0.1)
- Track consumption in real-time
- Implement hard stops when budget depletes
This forces teams to prioritize which queries matter most and prevents accidental privacy degradation through repeated access.
Key Takeaway: Epsilon controls privacy strength per operation, while privacy budgets enforce cumulative limits—lower epsilon means stronger privacy but less accuracy, and budgets prevent unlimited queries from eroding privacy over time.