EN
Back to the archive

The encyclopedia · R&D & Science · Technical decision · 2006–2014

Differential privacy made statistics safe by hiding each person

Dwork's definition bounds what one person can change, then masks it with calibrated noise — privacy you can prove.

Microsoft Research

The solution

By the mid-2000s, 'anonymized' datasets kept being re-identified, because removing names does not remove the fingerprints of the data itself. Cynthia Dwork's 2006 work reframed privacy as a property of the algorithm, not the dataset.

Differential privacy requires that adding or removing any single record change the output distribution only slightly. The Laplace mechanism implements it: compute the true answer, then add noise scaled to the query's ℓ1 sensitivity — the worst case one person can move the result.

The guarantee is formal and worst-case: no matter what auxiliary information an attacker has, one person's participation is barely visible. Composition theorems let a system answer many queries while tracking total privacy loss; the 2014 Dwork–Roth book systematized the theory.

Why it worked

  • The bound holds against arbitrary auxiliary knowledge
  • Laplace noise is calibrated to worst-case sensitivity
  • Guarantees compose across many queries
  • It protects individuals, not just datasets
What it achievedCalibrate noise to worst-case sensitivityinspired

What can be applied

Don't promise to hide data — promise one person's presence is indistinguishable. Measure the worst case one record can change the answer, then cover it with noise: a formal bound beats hope.

Aftermath

Differential privacy became the standard formal definition of private data analysis, adopted by statistical agencies and technology companies for official statistics and telemetry; the Dwork–Roth book remains its canonical reference.

Sources

spotted an error? The archive wants to know.

Related cases