Definition
Plain language
A strict guarantee that whatever a system publishes looks essentially the same whether or not any one person's data was included.
As stated in the literature
A formal privacy notion bounding the ratio of output probabilities between neighboring datasets by e^ε; the paper invokes it as the kind of distributional indistinguishability needed to close the leak.
Why it matters: It sets the bar for what a real confidentiality guarantee looks like — one that holds no matter how many clever queries an attacker asks — rather than merely hoping a system stays quiet about a secret.
For example, a survey report can say how many people in a town own a car while guaranteeing that no reader can tell whether any specific neighbor answered the survey at all.
Heard on the show
“Is this a deployment problem — normalize response lengths, resample, hand number generation to a tool — or does it have to be trained in as distributional invariance, the way differential privacy is?”Episode 246 — 160 Perfect Refusals, And The Refusals Were The Leak