Also known as: DP · privacy budget · epsilon
Differential privacy is a mathematical framework that adds carefully calibrated statistical noise to data or query results, providing a provable guarantee that the presence or absence of any single individual has negligible effect on the output.
Its strength is that the guarantee is formal rather than heuristic, and holds regardless of what auxiliary information an attacker possesses — which is precisely where traditional anonymisation fails.
The privacy budget, denoted epsilon, quantifies the trade-off: lower epsilon means stronger privacy and noisier results. Deployments frequently choose epsilon values high enough that the practical guarantee is much weaker than the framework's reputation implies, so the parameter matters as much as the label.
Anonymization
Anonymisation is the process of altering data so that individuals can no longer be identified from it, either directly or by combining it with other available information — as distinct from pseudonymisation, which only replaces identifiers with reversible tokens.
Data Minimization
Data minimisation is the principle that an organisation should collect and retain only the personal data actually necessary for a stated purpose, and no more, for no longer than needed.
Federated Learning
Federated learning is a technique for training a shared model across many devices without centralising the training data: each device computes updates locally, and only those model updates — not the underlying data — are sent to a server for aggregation.
Edge AI
Edge AI is the practice of running AI models directly on end-user devices such as phones and laptops rather than on remote servers, keeping input data local to the device.