Efficient heterogeneous federated learning method and system based on hybrid distillation, device, and medium

Patent metadata

Field	Value
Publication number	US-2026057246-A1
Application number	US-202519371654-A
Country	US
Kind code	A1
Filing date	Oct 28, 2025
Priority date	Jul 22, 2024
Publication date	Feb 26, 2026
Grant date	—

Abstract

An efficient heterogeneous federated learning method based on hybrid distillation includes: initializing, by a server, global model parameters, and setting a preset total number of training rounds and a number of clients participating in each of the training rounds; loading local datasets in the clients respectively, performing random transformations on the local datasets to generate client distillation data for the clients, sampling multiple sub-networks from an original network of each client, training each sub-network on the client distillation data to obtain updated local model parameters of each client, and uploading the updated local model parameters to the server; and receiving, by the server, the updated local model parameters, performing, by the server, server distillation based on the updated local model parameters and a preset auxiliary dataset to obtain updated global model parameters and an updated global model, and sending, by the server, the updated global model to the clients.

First claim

What is claimed is:

1 . An efficient heterogeneous federated learning method based on hybrid distillation, comprising:
step 1, initializing, by a server, global model parameters, and setting a preset total number of training rounds and a number of clients participating in each of the training rounds; step 2, loading local datasets in the clients respectively, performing random transformations on the local datasets to generate client distillation data for the clients, sampling a plurality of sub-networks from an original network of each of the clients, training each of the plurality of sub-networks based on the client distillation data of each of the clients to obtain updated local model parameters of each of the clients, and uploading the updated local model parameters to the server; step 3, receiving, by the server, the updated local model parameters of each of the clients, performing, by the server, server distillation based on the updated local model parameters of each of the clients and a preset auxiliary dataset to obtain updated global model parameters and an updated global model, and sending, by the server, the updated global model to the clients; and repeating step 2 to step 3 until the updated global model converges.

2 . The efficient heterogeneous federated learning method based on hybrid distillation as claimed in claim 1 , wherein the performing random transformations on the local datasets comprises:
performing scaling and rotating on the local datasets to obtain the client distillation data for the clients.

3 . The efficient heterogeneous federated learning method based on hybrid distillation as claimed in claim 1 , wherein the plurality of sub-networks have different network score widths.

4 . The efficient heterogeneous federated learning method based on hybrid distillation as claimed in claim 1 , wherein the training each of the plurality of sub-networks based on the client distillation data of each of the clients comprises:
calculating a Kullback-Leibler (KL) divergence between a softmax output of each of the plurality of sub-networks and an original softmax output of a local model of a corresponding one of the clients as a distillation loss, and dynamically assigning weights for each of the plurality of sub-networks based on prediction confidence of each of the plurality of sub-networks; and updating, based on the distillation loss and a traditional cross-entropy loss, local model parameters of each of the clients by using an optimization algorithm to obtain the updated local model parameters of each of the clients.

5 . The efficient heterogeneous federated learning method based on hybrid distillation as claimed in claim 1 , wherein the step 3 specifically comprises:
receiving, by the server, the updated local model parameters of each of the clients from the clients, and preforming, by the server, weight aggregation on the updated local model parameters of each of the clients to obtain a global model; and performing, based on the preset auxiliary dataset and by using a combination of soft prediction distillation and feature distillation, distillation on the global model to obtain the updated global model parameters and the updated global model, and sending the updated global model back to the clients.

6 . An efficient heterogeneous federated learning system based on hybrid distillation as claimed in claim 1 , comprising:
an initialization module, configured to make a server initialize global model parameters, and set a preset total number of training rounds and a number of clients participating in each of the training rounds; a client distillation module, configured to load local datasets in the clients respectively, perform random transformations on the local datasets to generate client distillation data for the clients, sample a plurality of sub-networks from an original network of each of the clients, train each of the plurality of sub-networks based on the client distillation data of each of the clients to obtain updated local model parameters of each of the clients, and upload the updated local model parameters to the server; and a server distillation module, configured for the server to receive the updated local model parameters of each of the clients, perform server distillation based on the updated local model parameters of each of the clients and a preset auxiliary dataset to obtain updated global model parameters and an updated global model, and send the updated global model to the clients; wherein the client distillation module and the server distillation module are further configured to repeat above steps until the updated global model converges.

7 . An electronic device, comprising: a memory and a processor, wherein the memory is configured to store a computer program, and the processor is configured to execute the computer program to make the electronic device implement the efficient heterogeneous federated learning method based on hybrid distillation as claimed in claim 1 .

8 . A computer-readable storage medium, wherein the computer-readable storage medium is stored with a computer program, and the computer program is configured to, when executed by a processor, implement the efficient heterogeneous federated learning method based on hybrid distillation as claimed in claim 1 .

Assignees

• Univ Dongguan Technology

Inventors

• Wang Yi
• Liu Senbin

Classifications

• G06N3/08

  Learning methods · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/048

  Activation functions · CPC title

• G06N3/098 Primary

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

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title