Protecting people in training data.
WHAT THE RISKS ARE
Models memorising training examples Inferring whether someone was in the training set Reconstructing training data from a model Sensitive attributes inferred from other data
WHAT MEMORISATION MEANS
A model reproducing specific training examples verbatim.
WHERE IT OCCURS
Large models trained on data containing unique strings.
WHAT REDUCES IT
Deduplication of training data Regularisation Limiting training on rare unique content
WHAT DIFFERENTIAL PRIVACY PROVIDES
A mathematical bound on what any individual's inclusion reveals.
WHAT IT COSTS
Accuracy, traded deliberately.
WHAT FEDERATED LEARNING DOES
Trains across devices without collecting the data centrally.
WHAT IT SUITS
Data that should not leave devices.
WHAT IT DOES NOT PROVIDE ALONE
Full privacy, since updates can leak information.
WHAT TO DO BEFORE TRAINING
Minimise what is included Remove direct identifiers Consider whether the data may lawfully be used this way
WHY THAT LAST POINT
Data collected for one purpose may not be usable for another.
WHAT TO PLAN
How to remove an individual's influence, if required.
WHAT TO ACCEPT
That retraining may be the only answer.
