← 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.