A federated learning system enabling selective model unlearning by reducing contribution weights of specific agents while maintaining global model integrity.

Patentradars sammanfattning

This invention relates to distributed machine learning systems where a global model is constructed from contributions of multiple agents using weighted disentangled parameters. Upon receiving an unlearn request from a specific agent, the system identifies that agent's contribution weight, reduces it, and updates the global model accordingly, then distributes the modified model to other agents. The approach enables targeted removal or reduction of individual agent contributions without retraining the entire system.

AI-genererad och redaktionellt processad.