Differential Privacy
A formal, mathematical guarantee that no individual record in a dataset can be detected from a system’s outputs.
Differential privacy adds carefully calibrated random noise to a computation (a query result, or gradient updates during training) so that the output’s probability distribution barely changes whether or not any single individual’s record was included — giving a provable bound, not just a best-effort promise, on how much any one person’s data could be inferred. It comes at a real cost in accuracy, tuned by a privacy budget parameter (ε), which is the core trade-off practitioners have to make.
Worked example: add calibrated random noise to data or query results so any single individual’s presence cannot be detected, giving a mathematical privacy guarantee controlled by a budget ε (epsilon) — smaller ε = more noise = more privacy. Gotcha: the privacy-utility tradeoff is real (small ε blurs real signal), and the budget COMPOSES — every query spends some ε, so a system answering unlimited queries eventually leaks; you must track and cap total spend.