Local vs Global Differential Privacy
What you'll learn: You'll discover the critical difference between trusting a central curator to add noise (global DP) versus having each user add noise to their own data before sharing (local DP).
The Trust Divide
Differential privacy can be deployed in two fundamentally different trust models. Understanding which to use depends entirely on who you're willing to trust with your raw data.
Global Differential Privacy (Centralized Model)
In global DP, users send their true data to a trusted central server (the "curator"). This server sees everyone's real information, applies differential privacy mechanisms (like Laplace or Gaussian noise), and releases only the noised results.
Analogy: You submit your real medical data to a hospital research database. The hospital promises to add noise before publishing statistics. You must trust the hospital completely—they see everything.
Advantages:
- Much better accuracy (lower noise) for the same privacy budget ε
- Enables complex queries and correlations
- The curator can optimize noise across queries
Local Differential Privacy (Distributed Model)
In local DP, each user adds noise to their own data before sending anything to the server. The server never sees true data—only pre-noised values from individuals.
Analogy: Before filling out a survey, you flip coins to randomize your answers, then submit the garbled version. The survey company never knows your real answers, only noisy ones.
Advantages:
- No need to trust the server at all
- Protects against server breaches and insider threats
Trade-off: Local DP requires significantly more noise per individual to achieve the same privacy guarantee, often reducing utility dramatically.
Which to Choose?
Use global DP when a trusted curator exists. Use local DP when no entity should ever see raw data—even temporarily.
Key Takeaway: Global DP trusts a central server with raw data and offers better accuracy; local DP eliminates trust requirements by having users add noise themselves, but at the cost of much higher noise and lower utility for the same privacy level.