Verfahren, System und Computerprogramm zum Ausgleich von Privatsphäre, Nutzen und Fairness bei Föderiertem Lernen

Details

Veröffentlichungs-Nr.

EP4730208

Anmeldetag

17. Oktober 2024

Veröffentlichung

22. April 2026

Rechtsraum

EP

IPC

G06N3/098, G06N3/084, G06N3/09

Abstract

A method, system and computer program to balance privacy, utility, and fairness in Federated Learning are provided. The method comprises selecting, by a central server, for a first training round, a plurality of available computing devices to train a ML model using a FL strategy; establishing, by each device, a target disparity value the ML model should fulfill at the end of the training; training, by each device, the ML model by: selecting, from a dataset, a batch of training data for training the ML model; computing a first loss; computing a second loss; computing a global loss function by summing the first and second loss and balancing them with a λ parameter, the latter being dynamically updated; sharing, by each device, the trained ML model with the central server, which further aggregates them; and repeating the method for a next training round until achieving a predefined number of training rounds.

Anmelder

Firma

Telefónica Innovación Digital, S.L.

Land

🇪🇸 Spanien