Method and apparatus for performing horizontal federated learning

Patent metadata

Field	Value
Publication number	US-12488140-B2
Application number	US-202217939665-A
Country	US
Kind code	B2
Filing date	Sep 7, 2022
Priority date	Sep 7, 2021
Publication date	Dec 2, 2025
Grant date	Dec 2, 2025

Abstract

The present disclosure discloses a method for federated learning, including: requesting, by a target application, a user to provide a federated learning data access right to access privacy data; calling, by the target application, a unified federated learning service (UnifiedFED) application to perform horizontal federated learning, acquiring, by the UnifiedFED application, the privacy data isolated from the target application according to the federated learning data access right, receiving, by the UnifiedFED application, non-privacy data from the target application, and providing, by the UnifiedFED application, the privacy data and the non-privacy data to an artificial intelligence (AI) model for federated learning training.

First claim

What is claimed is:

1 . A method for performing federated learning, comprising:
requesting, by a target application, a user to provide a federated learning data access right to access privacy data, wherein the federated learning data access right is a right set specific to the federated learning, under the federated learning data access right, a privacy data access and an artificial intelligence (AI) model interaction are completed by a unified federated learning service (UnifiedFED) application and the target application cannot acquire a content of the privacy data; calling, by the target application, the UnifiedFED application to perform federated learning, acquiring, by the UnifiedFED application, the privacy data isolated from the target application according to the federated learning data access right only for the federated learning, receiving, by the UnifiedFED application, non-privacy data from the target application, and providing, by the UnifiedFED application, the privacy data and the non-privacy data to the AI model for federated learning training, wherein the privacy data is accessed only by the UnifiedFED application, and not by the target application requesting access to the privacy data and directly transmitted to the UnifiedFED application without going through the target application.

2 . The method as claimed in claim 1 , wherein the acquiring of the privacy data comprises:
in response to the target application about the federated learning data access right, identifying a type of the federated learning data access right; and acquiring the privacy data based on the identified type.

3 . The method as claimed in claim 2 , the method further comprises:
identifying the type of the federated learning data access right as being allowed to have the federated learning data access right only at a current time; and acquiring the privacy data based on the identified type.

4 . The method as claimed in claim 2 , the method further comprises:
identifying the type of the federated learning data access right as being allowed to have the federated learning data access at all times, authorizing the federated learning data access right to the target application, and acquiring the privacy data based on the identified type.

5 . The method as claimed in claim 2 , the method further comprises:
identifying the type of the federated learning data access right as not being allowed to have the federated learning data access at all times; and rejecting the target application to obtain the federated learning data access right.

6 . The method as claimed in claim 1 , wherein the method further comprises:
acquiring, by the UnifiedFED application, self-owned data of the target application, and preprocessing the self-owned data based on a user setting and a parameter that is transmitted when the target application calls the UnifiedFED application to acquire the preprocessed data that is input into the AI model.

7 . The method according to claim 1 , wherein the method further comprises:
determining, by the UnifiedFED application, schedules of the federated learning training as executing a task of the federated learning training immediately; when a current device status meets a condition notifying the target application that model training can be performed, notifying a cloud server by the target application that the model training can be performed on a local device, and downloading the model when the local device is chosen by the cloud server or in advance; and when the current device status does not meet the condition notifying the target application that a request for the federated learning data access right is rejected, and notifying, by the target application, the cloud server to choose another device to perform the model training.

8 . The method as claimed in claim 1 , wherein the method further comprises:
determining, by the UnifiedFED application, schedules of the federated learning training as executing a task of the federated learning training asynchronously; periodically checking a current device status, when the current device status meets a condition, notifying the target application that the model training can be performed, notifying a cloud server by the target application that the model training can be performed on a local device, and downloading the model when the local device is chosen by the cloud server or in advance.

9 . The method as claimed in claim 7 , wherein the UnifiedFED application determines if the current device status meets the condition based on at least one of a task type, a model size, a number of neural network parameters to be trained, a service condition of a processor, a battery, or a network of a device that performs the method, a data volume to be trained, and a user setting,
wherein the task type comprises a predicting task and a training task, with the predicting task having a priority level higher than that of the training task; and the model size, the number of neural network parameters to be trained, and the data volume to be trained are used to estimate a time required for training.

10 . The method as claimed in claim 1 , further comprises:
transmitting, by the UnifiedFED application, a gradient to a cloud server; and receiving, by the UnifiedFED application, updated AI model from the cloud server.

11 . An apparatus for federated learning, the apparatus comprising:
a memory storing one or more instructions; and at least one processor configured to execute the one or more instructions to: determine, when receiving scheduling from a target application, if the target application has a federated learning data access right to access privacy data, wherein the federated learning data access right is a right set specific to the federated learning, under the federated learning data access right, a privacy data access and an artificial intelligence (AI) model interaction are completed by a unified federated learning service (UnifiedFED) application and the target application cannot acquire a content of the privacy data; and in response to determining that the target application has the federated learning data access right to access the privacy data, acquire the privacy data that is isolated from the target application only for the federated learning, receive non-privacy data from the target application, and provide the privacy data and the non-privacy data to the AI model to perform federated learning training, wherein the privacy data is acquired only by the UnifiedFED application, and not by the target application having the federated learning data access right and directly transmitted to the UnifiedFED application without going through the target application.

12 . The apparatus as claimed in claim 11 , wherein the at least one processor is further configured to execute the one or more instructions to:
in response to the target application about the federated learning data access right, identify a type of the federated learning data access right; and acquire the privacy data based on the identified type.

13 . The apparatus as claimed in claim 11 , wherein the at least one processor is further configured to preprocess user data based on a user configuration and a parameter that are transmitted when the target application calls the UnifiedFED application to acquire the preprocessed data that is input into the AI model, and the user data comprises the privacy data and self-owned data of the target application.

14 . The apparatus as claimed in claim 11 , wherein the at least one processor is further configured to:
determine schedules of the federa

Assignees

• Samsung Electronics Co Ltd

Inventors

• Wang Chengjun

Classifications

• G06N20/20

  Ensemble learning · CPC title

• G06N3/098

  Distributed learning, e.g. federated learning · CPC title

• G06F21/6245 Primary

  Protecting personal data, e.g. for financial or medical purposes · CPC title