Adaptive Modellzuweisung und Neuzuweisung Beim Vertikalen Föderierten Lernen

Details

Veröffentlichungs-Nr.

EP4790557

Anmeldetag

5. Februar 2025

Veröffentlichung

12. August 2026

Rechtsraum

EP

IPC

G06F9/50, G06N3/098

Abstract

The disclosure relates to a method performed by a vertical federated learning, VFL server (10, 10a), the method comprising: obtaining (31, 31a, 31b) client information for a plurality of VFL clients (11, 12, 13, 14, 11a, 12a, 13a, 14a), the client information for a VFL client including data indicative of a reliability and/or a resource availability of the VFL client; transmitting (34, 34a, 34b) allocation instructions, the allocation instructions defining associations of local models to respective VFL clients, wherein the allocation instructions take account of the client information; obtaining (35, 35a, 35b) an update to the client information for a VFL client during a use of a local model by the VFL client in a training task (22) or an inference task (23); transmitting (36, 36a, 36b) at least one updated allocation instruction, an updated allocation instruction defining a reassociation of a local model to a respective VFL client, and wherein the at least one updated allocation instruction takes account of the update to the client information; and transmitting (33, 33a, 33b) model information for at least one of the local models, the model information comprising one or more of a model importance, a model size, and a model complexity, wherein the allocation instructions and/or the at least one updated allocation instruction take account of the model information.

Anmelder

Firmen

Mitsubishi Electric R&D Centre Europe B.V.
MITSUBISHI ELECTRIC CORPORATION

Land

🇳🇱 Niederlande

Fachgebiete