Skip to content
Vol. 1 · Ed. 2026
CyberGlossary
Entry № 317

Differential Privacy

What is Differential Privacy?

Differential PrivacyA mathematical framework that quantifies privacy loss when releasing statistics or training models, by adding calibrated noise so any single individual's contribution is provably bounded.


Differential privacy, formalized by Dwork, McSherry, Nissim, and Smith, guarantees that the probability of any output changes by at most a small factor e^epsilon (and optionally delta) when one record is added or removed from a dataset. Mechanisms include the Laplace, Gaussian, and exponential mechanisms, as well as DP-SGD for machine learning. Cumulative privacy loss is tracked with a privacy budget (epsilon-delta) and advanced composition or moments accountants. The U.S. Census Bureau (2020 decennial), Apple, Google, and Microsoft have deployed it for telemetry and statistics. Unlike syntactic models (k-anonymity, l-diversity), it provides provable, future-proof guarantees regardless of adversary auxiliary knowledge.

Examples

  1. 01

    Apple's keyboard suggestions reporting emoji frequencies via local differential privacy.

  2. 02

    Training a healthcare model with DP-SGD so individual patient records cannot be memorized.

Frequently asked questions

What is Differential Privacy?

A mathematical framework that quantifies privacy loss when releasing statistics or training models, by adding calibrated noise so any single individual's contribution is provably bounded. It belongs to the Privacy & Data Protection category of cybersecurity.

What does Differential Privacy mean?

A mathematical framework that quantifies privacy loss when releasing statistics or training models, by adding calibrated noise so any single individual's contribution is provably bounded.

How does Differential Privacy work?

Differential privacy, formalized by Dwork, McSherry, Nissim, and Smith, guarantees that the probability of any output changes by at most a small factor e^epsilon (and optionally delta) when one record is added or removed from a dataset. Mechanisms include the Laplace, Gaussian, and exponential mechanisms, as well as DP-SGD for machine learning. Cumulative privacy loss is tracked with a privacy budget (epsilon-delta) and advanced composition or moments accountants. The U.S. Census Bureau (2020 decennial), Apple, Google, and Microsoft have deployed it for telemetry and statistics. Unlike syntactic models (k-anonymity, l-diversity), it provides provable, future-proof guarantees regardless of adversary auxiliary knowledge.

How do you defend against Differential Privacy?

Defences for Differential Privacy typically combine technical controls and operational practices, as detailed in the full definition above.

What are other names for Differential Privacy?

Common alternative names include: DP, Epsilon-Differential Privacy.

Related terms