Differential Privacy

What is differential privacy?

Say you want to know the average salary in a room, but nobody will tell you their number. Differential privacy adds a little carefully designed noise to the answer, so whoever sees the result can't tell whether any specific person is in the data — or what their number was. The word “differential” comes from comparing two datasets that differ by just one person and seeing how far the outputs drift. Keep that difference small enough and any single individual disappears. It isn't hiding the data; it's making each person's contribution indistinguishable.

How can something be both public and private?

The noise is tunable
How much noise to add is a precise math problem. More noise means more privacy but fuzzier results, so you pick a trade-off.
It gives you a privacy budget
Every query spends a bit of that budget. Spend too much and the system starts refusing to answer, so nobody can hammer the same question and stitch the truth back together.

Where is it used?

Apple and Google use it to learn from typing habits and browsing data without exposing individuals. The US Census uses a similar idea to publish population stats. And now large-model training uses it too — the model “sees” the data without memorizing any one person.

The core problem it solves

Old-school anonymization relied on deleting names and ID numbers, but attackers can still re-identify people from indirect clues — zip code, age, a movie you liked. Differential privacy closes that door mathematically: no matter how clever the attacker, they can't tell if you're in there.

Bottom line: differential privacy mixes a little noise into the data, keeping the stats useful while making any single person impossible to trace.