Techniques allowing a model to unlearn specific data without complete retraining, but without guarantee of a total or verifiable forgetting.
Machine unlearning gathers the techniques allowing a learning model to unlearn specific data by modifying its internal parameters, without undergoing the complete retraining that would be computationally prohibitive. Born about a decade ago, the discipline has produced so-called certified unlearning frameworks, which aim to guarantee mathematically, through probabilistic bounds, that the influence of the removed data has been sufficiently reduced. Its limits remain serious, for there exists no universally recognised method to verify that a model has effectively forgotten, and the entanglement of an individual's data with that of thousands of others makes their removal without degradation very delicate. Machine unlearning therefore offers a real but partial erasure, approximate and hard to prove, far from the clean suppression that the right to be forgotten postulates.
To answer an erasure request, a vendor applies a machine-unlearning technique to its model, without however being able to demonstrate to the regulator that all trace of the person has disappeared.
machine unlearning, désapprentissage automatique, oubli certifié