Privacy preserving cooperative learning in untrusted environments

Patent metadata

Field	Value
Publication number	US-12229280-B2
Application number	US-202217695820-A
Country	US
Kind code	B2
Filing date	Mar 15, 2022
Priority date	Mar 16, 2021
Publication date	Feb 18, 2025
Grant date	Feb 18, 2025

Abstract

Aspects of the present disclosure provide systems, methods, and computer-readable storage media that support cooperative training of machine learning (ML) models that preserves privacy in untrusted environments. For example, a server (or cloud-based computing device(s)) may be configured to “split” an initial ML model into various partial ML models, some of which are provided to client devices for training based on client-specific data. Output data generated during the training at the client devices may be provided to the server for use in training corresponding server-side partial ML models. After training of the partial ML models is complete, the server may aggregate the trained partial ML models to construct an aggregate ML model for deployment to the client devices. Because the client data is not shared with other entities, privacy is maintained, and the splitting of the ML models enables offloading of computing resource-intensive training from client devices to the server.

First claim

What is claimed is:

1. A method for cooperative training of machine learning models, the method comprising:
generating, by one or more processors, a first parameter set corresponding to a first partial machine learning (ML) model, a second parameter set corresponding to a second partial ML model, a third parameter set corresponding to a third partial ML model, and a fourth parameter set corresponding to a fourth partial ML model,
wherein the first parameter set and the third parameter set correspond to a first splitting of an initial ML model design, and the second parameter set and the fourth parameter set correspond to a second splitting of the initial ML model design;

initiating, by the one or more processors, transmission of the first parameter set to a first client device and of the second parameter set to a second client device;
modifying, by the one or more processors, the third parameter set based on first output data received from the first client device,
wherein the first output data represents output of a first trained ML model that is based on the first parameter set and trained using first client data;

modifying, by the one or more processors, the fourth parameter set based on second output data received from the second client device,
wherein the second output data represents output of a second trained ML model that is based on the second parameter set and trained using second client data; and

aggregating, by the one or more processors, at least the modified third parameter set and the modified fourth parameter set to create an aggregate parameter set corresponding to an aggregate ML model.

2. The method of claim 1 , wherein:
a structure of the first partial ML model is different from a structure of the second partial ML model; and
the structure of the first partial ML model and the structure of the second partial ML model include a number of layers associated with the respective partial ML model, a number of nodes per layer associated with the respective partial ML model, or a combination thereof.

3. The method of claim 1 , wherein a structure of the first partial ML model is the same as a structure of the second partial ML model.

4. The method of claim 1 , wherein:
the first splitting of the initial ML model design is based on first privacy leakage and preservation corresponding to the first client device; and
the second splitting of the initial ML model design is based on second privacy leakage and preservation corresponding to the second client device.

5. The method of claim 1 , wherein a structure of the third partial ML model is different than a structure of the first partial ML model, a structure of the second partial ML model is different than a structure of the fourth partial ML model, or a combination thereof.

6. The method of claim 1 , wherein modifying the third parameter set comprises:
providing, by the one or more processors, the first output data as training data to the third partial ML model.

7. The method of claim 6 , further comprising:
determining, by the one or more processors, first gradient data based on output of the third partial ML model during training; and
initiating, by the one or more processors, transmission of the first gradient data to the first client device.

8. The method of claim 6 , further comprising:
receiving, by the one or more processors, additional output data from the first client device,
wherein the additional output data represents output of the first trained ML model based on the first gradient data; and
providing, by the one or more processors, the additional output data as further training data to the third partial ML model.

9. The method of claim 6 , wherein modifying the fourth parameter set comprises:
providing, by the one or more processors, the second output data as training data to the fourth partial ML model.

10. The method of claim 9 , further comprising:
initiating, by the one or more processors, transmission of second gradient data to the second client device,
wherein the second gradient data is based on output of the fourth partial ML model during training; and

providing, by the one or more processors, second additional output data as further training data to the fourth partial ML model,
wherein the second additional output data is received from the second client device and representing output of the second trained ML model based on the second gradient data.

11. The method of claim 1 , wherein aggregating the modified third parameter set and the modified fourth parameter set comprises averaging one or more structural feature values corresponding to the modified third parameter set and one or more structural parameter values corresponding to the modified fourth parameter set.

12. The method of claim 1 , wherein aggregating the modified third parameter set and the modified fourth parameter set comprises:
weighting one or more structural feature values corresponding to the modified third parameter set;
weighting one or more structural parameter values corresponding to the modified fourth parameter set; and
averaging the one or more weighted structural feature values corresponding to the modified third parameter set and the one or more weighted structural feature values corresponding to the modified fourth parameter set.

13. The method of claim 12 , wherein weights associated with the one or more weighted structural feature values corresponding to the modified third parameter set are based on a data size of the first client data, an amount of resources associated with the first client device, a priority associated with the first client device, or a combination thereof.

14. The method of claim 1 , further comprising initiating, by the one or more processors, deployment of the aggregate parameter set to one or more client devices for creation of one or more ML models at the one or more client devices.

15. The method of claim 1 , further comprising:
obtaining, by the one or more processors, input data corresponding to a task to be performed by an ML model corresponding to the aggregate parameter set;
providing, by the one or more processors, the input data to the ML model to generate a predicted output; and
initiating, by the one or more processors, performance of one or more actions based on the predicted output.

16. A system for cooperative training of machine learning models, the system comprising:
a memory; and
one or more processors communicatively coupled to the memory, the one or more processors configured to:
generate a first parameter set corresponding to a first partial machine learning (ML) model, a second parameter set corresponding to a second partial ML model, a third parameter set corresponding to a third partial ML model, and a fourth parameter set corresponding to a fourth partial ML model,
wherein the first parameter set and the third parameter set correspond to a first splitting of an initial ML model design, and the second parameter set and the fourth parameter set correspond to a second splitting of the initial ML model design;

initiate transmission of the first parameter set to a first client device and of the second parameter set to a second client device;
modify the third parameter set based on first output data received from the first client device,
wherein the first output data represents output of a first trained ML model that is based on the first parameter set and trained using first client data;

modify the fourth parameter set based on second output data received from the second client device,
wherein the second output data represents output of

Assignees

• Accenture Global Solutions Ltd

Inventors

• Ding aolin
• Hassanzadeh Amin

Classifications

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/082

  modifying the architecture, e.g. adding, deleting or silencing nodes or connections · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/08

  Learning methods · CPC title