Differential Privacy
A mathematical guarantee that a data analysis reveals almost nothing about any single individual — privacy you can actually prove, not just promise.
Reviewed July 16, 2026Stable
When not to use it
- When you need exact answers and can't tolerate any noise — differential privacy always trades accuracy for privacy.
- When the dataset is tiny — the noise needed for a real guarantee can swamp the signal.
- As a reason to skip other protections — it bounds statistical leakage, not access control or breaches.
Reach for something else instead
- k-anonymity and related methods — older, weaker, and defeated by auxiliary information (which is what DP fixes).
- Federated learning — keeps raw data local (often combined with DP rather than replacing it).
- Secure computation / homomorphic encryption for a different privacy problem (computing on encrypted data).
Sources & further reading
- Dwork et al. (2006), Calibrating Noise to Sensitivity in Private Data Analysis — the founding paper.
- Dwork & Roth (2014), The Algorithmic Foundations of Differential Privacy — the standard reference.
- Abadi et al. (2016), Deep Learning with Differential Privacy — DP-SGD, private training of neural networks.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Treating the privacy budget as unlimited — every query spends epsilon, and it runs out.
- Choosing epsilon without understanding it — a large epsilon offers little real protection.
- Assuming it protects against everything — it bounds inference from outputs, not breaches or misuse of access.
At a glance
FieldSafety & Ethics
Guaranteeprovable per-individual privacy
Knobepsilon (privacy budget)
Methodcalibrated noise
DifficultyAdvanced