How can useful statistics reveal less about any one person?

A group's published total can reveal one person's response when someone already knows everyone else's answers. Subtracting the known responses exposes the remaining one, even if names were removed. Differential privacy addresses this disclosure problem by limiting how much one person's contribution can change the probabilities of a randomized release.

The comparison uses neighboring datasets, defined to differ in a specified individual contribution. A randomized mechanism must keep the probabilities of possible outputs sufficiently similar across those datasets, according to stated privacy parameters. It constrains the release process rather than declaring one particular published number harmless.

For a count in which each person contributes at most one, adding or removing one person can change the exact answer by at most one. A suitable mechanism adds random noise scaled to that sensitivity and to the desired privacy level. A released value can then be compatible with either neighboring dataset, making one person’s participation harder to distinguish from the output.

The protection is about bounded inference, not a promise that nobody can guess anything. Someone may already know facts from other sources, and the aggregate can still reveal real patterns about a population. More protective settings generally require more uncertainty for the same query. That uncertainty can be small compared with a large group count yet large relative to a tiny subgroup.

Repeated releases accumulate information, so their combined privacy cost needs accounting through a privacy budget. Contribution limits, the neighboring-dataset definition and correct software also matter. The overlapping curves illustrate reduced distinguishability without supplying a numerical guarantee: overlap alone cannot prove differential privacy, whose protection belongs to the specified release mechanism and its parameters.