The usual technique is to add carefully calibrated random noise. Suppose a hospital publishes how many patients in a county have diabetes. If the exact count changes by one when a single person joins the dataset, a determined attacker comparing releases could learn that person’s diagnosis. With differential privacy, the published number has noise added so the answer looks essentially the same with or without any one person, while staying accurate in aggregate. A setting usually called epsilon, the “privacy budget,” controls the trade-off: less noise means more accuracy and less privacy, and every additional query spends some of the budget.
In episode 22 at 38:58, Databricks’ Denny Lee explains “differential privacy in its most simplistic form” as privacy-preserving histograms for aggregate queries, using a diabetes-count example, after tracing the idea back to research on re-identifying people from supposedly anonymous data. Galileo sponsored the show at the time; Conor mentions the company in the episode at this point.
The best-known large deployment is the U.S. Census Bureau’s use of a differential-privacy framework for 2020 Census data; as of October 24, 2026, its page says it is evaluating alternatives after a Commerce Department order prohibiting noise infusion. For AI teams, the technique also appears in training (to limit what a model memorizes about individuals) and in analytics on user data. Editorially, it complements rather than replaces PII redaction: redaction removes obvious identifiers, while differential privacy limits what can be inferred from what remains. It also bears on synthetic data, which is not automatically private.
Hear it from the guest
“Differential privacy in its most simplistic form in terms of privacy preserving histograms”
Quotes lightly edited to remove filler words.
Sources
- Dwork and Roth: The Algorithmic Foundations of Differential Privacy — Author-hosted monograph by the field’s founders setting out the formal definition, the promise of differential privacy and its limits; checked October 24, 2026.
- NIST SP 800-226: Guidelines for Evaluating Differential Privacy Guarantees — March 2025 NIST guidance on understanding and evaluating differential privacy claims; checked October 24, 2026.
- U.S. Census Bureau: Understanding Differential Privacy — Says 2020 Census data used a protection framework based on differential privacy, and that the Bureau is now evaluating alternatives to comply with a Commerce Department order prohibiting noise infusion; checked October 24, 2026.
