Identity forgetting without the original forget set: delete the prototype, then readapt on limited support

Identity-level forgetting can be carried out by editing an external memory of prototypes instead of re-exposing the model to the complete original forget set at request time. Forgetting requests are usually handled with the target identity's full training data available at request time, whereas this setting covers requests where the complete original forget set is absent. On the identity partitioning and one-shot support construction of 1-SHUI, MCU reaches a favorable empirical trade-off among model utility, membership distinguishability, and request-time computational efficiency relative to the evaluated baselines.

A controllable external identity-prototype memory sits beside the host model, so forgetting becomes deleting a memory entry and then readapting the model. The memory is a Modern Hopfield Network that stores identity prototypes formed by aggregating the feature representations associated with each training identity. A Prototype Adaptation Network transforms the retrieved memory representation and a Dependency Gating Unit controls its contribution to downstream prediction; forgetting removes the corresponding prototype and then readapts with limited support information for the target identities together with retained data.

The work targets settings where only limited target-identity information is available when a forgetting request is processed and the complete original forget set is absent from request-time gradient-based readaptation, while retained data may still be used to preserve downstream-task utility. Evaluation rests on the identity partitioning and one-shot support construction of 1-SHUI, compared against the evaluated baselines. The authors position the results as empirical evidence for the potential of controllable external memory in machine unlearning, not as a formal guarantee of complete information removal.

The authors state that data-absent here specifically means the absence of the complete original forget set during request-time gradient-based readaptation, and that retained data may still be used. Results are framed as empirical evidence rather than a formal guarantee of complete information removal, so behavior on measures such as membership distinguishability should be read within that evaluation setting.