← Entropy Atlas

Secret & unpredictable

Differential-privacy noise

Carefully calibrated random noise limits what a published result reveals about one person.

Evaluation card

What “plenty” means here.

Required property
The specified noise distribution, protected generation, and privacy accounting across repeated releases.
If it fails
Biased, truncated, predictable, or reused noise can invalidate the promised privacy bound.
Relevant attacker
An analyst combining releases with auxiliary data to infer an individual contribution.
The Plenty Line
Implement a reviewed mechanism, sample securely when the model assumes secret noise, and track the total privacy budget.
Why more is not automatically better
More noise can destroy utility; privacy depends on calibrated scale and composition, not maximal randomness.

Where it appears

Recognize the pattern.

  • 01private statistics
  • 02telemetry aggregation
  • 03privacy-preserving machine learning

The Plentropy rule

Protect the requirement, then stop.

Implement a reviewed mechanism, sample securely when the model assumes secret noise, and track the total privacy budget.
Once that claim has comfortable evidence and margin, improve the next limiting factor instead of worshipping a larger entropy number.