Adding calibrated noise to protect individuals
Differential privacy is a mathematical framework for limiting how much any single data point can influence, and therefore be inferred from, a model's output. In practice, this means adding carefully calibrated statistical noise, either to the data, the training process, or the model updates themselves, so that the presence or absence of any one individual's data doesn't meaningfully change what an observer can learn.
The core guarantee is comparative: with differential privacy applied, an outside observer looking at the output can't tell with confidence whether any particular person's data was included at all. This makes it a natural complement to federated learning, since it directly targets the residual leakage that model updates can carry, adding a layer of protection to the updates themselves rather than just controlling where raw data travels.