Differential Privacy for Neural Data

A formal privacy guarantee, controlled by a budget epsilon, that bounds how much any single individual's data can change what an analysis or trained model reveals; a smaller epsilon means stronger privacy and more added noise. Applied to neural data it underlies private training (for example differentially private stochastic gradient descent) and private release of statistics from EEG, fMRI, or spike datasets.

Neural data stresses the framework: it is high-dimensional and strongly correlated in both time and space, so the unit of privacy (one sample, one session, one person) must be chosen carefully, and naive per-sample accounting under-protects a person who contributes many correlated samples. The noise needed for a meaningful epsilon can markedly reduce decoding accuracy, so most work reports a utility-privacy tradeoff rather than a free lunch; local differential privacy (noise added on-device before sharing) protects against an untrusted aggregator but costs even more utility.