Differential privacy

Differential privacy is a mathematical definition of privacy under which the result of an analysis changes very little whether or not any one person's data is included, achieved by adding carefully calibrated random noise.

NIST's guidelines, final since March 2025, describe it as a framework that quantifies the privacy loss to people when their data appears in a dataset (NIST SP 800-226, March 2025). Unlike masking or aggregation alone, it gives a provable bound on how much any released result can reveal about one person.

Two ideas drive every implementation. The parameter epsilon sets the strength: a smaller epsilon means more noise, stronger privacy and less accurate results. The privacy budget limits how many queries can run, because each answer leaks a little and the leaks add up (IAB Tech Lab, Differential Privacy Guidance). In advertising, clean rooms and measurement services apply it to counts, sums and conversion reports. The IAB Tech Lab suggests asking vendors where noise is applied, which epsilon values they use and how they manage the budget.

Example: A clean room reports 1,013 converters for a campaign where the true count is 1,000. The noise hides whether any single customer is in the data, while the advertiser's decision does not change. A report on 12 converters would carry the same noise and be far less useful.

Related terms: k-anonymity, clean room, ADMaP, PAIR, hashed email.

Authoritative source: NIST, SP 800-226, "Guidelines for Evaluating Differential Privacy Guarantees" (March 2025).

Read the full guide: AWS Clean Rooms vs Snowflake vs LiveRamp.