Federated machine learning management

Patent metadata

Field	Value
Publication number	US-12380361-B2
Application number	US-202117357626-A
Country	US
Kind code	B2
Filing date	Jun 24, 2021
Priority date	Jun 24, 2021
Publication date	Aug 5, 2025
Grant date	Aug 5, 2025

Abstract

Techniques are disclosed in which a computer system receives, from a plurality of user computing devices, a plurality of device-trained models and obfuscated sets of user data stored at the plurality of user computing devices, where the device-trained models are trained at respective ones of the plurality of user computing devices using respective sets of user data prior to obfuscation. In some embodiments, the server computer system determines similarity scores for the plurality of device-trained models, wherein the similarity scores are determined based on a performance of the device-trained models. In some embodiments, the server computer system identifies, based on the similarity scores, at least one of the plurality of device-trained models as a low-performance model. In some embodiments, the server computer system transmits, to the user computing device corresponding to the low-performance model, an updated model.

First claim

What is claimed is:

1. A method, comprising:
receiving, by a server computer system from a plurality of user computing devices, a plurality of device-trained models and obfuscated sets of user data stored at the plurality of user computing devices, wherein the device-trained models are trained at respective ones of the plurality of user computing devices using respective sets of user data prior to obfuscation;
determining, by the server computer system using a machine learning algorithm and based on characteristics specified in the obfuscated sets of user data, similarity scores for the plurality of device-trained models, wherein the similarity scores indicate a similarity between two or more of the plurality of device-trained models;
executing, by the server computer system based on the similarity scores indicating that two or more of the plurality of device-trained models are similar, the two or more device-trained models to generate risk scores for the obfuscated sets of user data;
identifying, by the server computer system based on the risk scores output by the two or more device-trained models, one of the two or more device-trained models as a low-performance model relative to the two or more device-trained models;
in response to identifying the low-performance model, transmitting, by the server computer system to a particular user computing device corresponding to the low-performance model, an updated model to replace the low-performance model at the particular user computing device; and
receiving, from the particular user computing device, a risk score generated using the updated model for a user request received at the particular user computing device.

2. The method of claim 1 , wherein the particular user computing device is a first user computing device, wherein the risk score received from the particular user computing device is a first risk score, and wherein the method further comprises:
receiving, by the server computer system, from a second user computing device of the plurality of user computing devices, a second risk score for a user request received at the second user computing device, wherein the second risk score is generated by the second user computing device using a device-trained model of the plurality of device-trained models; and
determining, by the server computer system based on a plurality of rules associated with the user request, a decision for the user request; and
transmitting, by the server computer system to the second user computing device, the decision for the user request.

3. The method of claim 2 , wherein the plurality of rules associated with the user request are selected based on one or more characteristics of the following types of characteristics: a location of the user computing device, a type of user request received at the user computing device, and one or more entities indicated in the user request.

4. The method of claim 1 , further comprising:
generating, by the server computer system prior to the transmitting, the updated model, wherein generating the updated model includes generating an aggregated model by combining two or more of the plurality of device-trained models received from the plurality of user computing devices.

5. The method of claim 4 , wherein generating the updated model further includes:
inputting the obfuscated set of user data received from a user computing device corresponding to the low-performance model into the aggregated model.

6. The method of claim 1 , wherein the machine learning algorithm is a clustering algorithm, and wherein the executing is performed based on:
identifying, based on the similarity scores, that the two or more device-trained models are nearest neighbors.

7. The method of claim 1 , further comprising:
storing, by the server computer system in a database based on domain information and geographic region information, the updated model.

8. The method of claim 7 , further comprising:
receiving, by the server computer system from a new user computing device that newly downloaded an application of the server computer system, a request for a machine learning model;
selecting, by the server computer system from the database, an updated model, wherein the selecting is performed based on searching the database according to one or more characteristics in an obfuscated set of user data received from the new user computing device; and
transmitting, by the server computer system to the new user computing device, the selected updated model.

9. A non-transitory computer-readable medium having instructions stored thereon that are executable by a server computer system to perform operations comprising:
receiving, from a plurality of user computing devices, a plurality of device-trained models and obfuscated sets of user data stored at the plurality of user computing devices, wherein the device-trained models are trained at respective ones of the plurality of user computing devices using respective sets of user data prior to obfuscation;
determining, using a clustering algorithm and based on characteristics specified in the obfuscated sets of user data, similarity scores for the plurality of device-trained models, wherein the similarity scores indicate a similarity between two or more of the plurality of device-trained models;
executing, based on the similarity scores indicating that two or more of the plurality of device-trained models are similar, the two or more device-trained models to generate risk scores for the obfuscated sets of user data;
identifying, based on the risk scores output by the two or more device-trained models, one of the two or more device-trained models as a low-performance model relative to the two or more device-trained models;
in response to identifying the low-performance model, transmitting to a particular user computing device corresponding to the low-performance model, an updated model to replace the low-performance model at the particular user computing device; and
receiving, from the particular user computing device, a risk score generated using the updated model for a user request received at the particular user computing device.

10. The non-transitory computer-readable medium of claim 9 , wherein the particular user computing device is a first user computing device, wherein the risk score received from the particular user computing device is a first risk score, and wherein the operations further comprise:
receiving, from a second user computing device of the plurality of user computing devices, a second risk score for a user request received at the second user computing device, wherein the second risk score is generated by the second user computing device using a device-trained model; and
determining, based on a plurality of rules associated with the user request, a decision for the user request; and
transmitting, to the second user computing device, the decision for the user request.

11. The non-transitory computer-readable medium of claim 10 , wherein the plurality of rules associated with the user request are selected based on one or more characteristics of the following types of characteristics: a location of the second user computing device, a type of user request received at the second user computing device, and one or more entities indicated in the user request.

12. The non-transitory computer-readable medium of claim 9 , wherein one of the plurality of device-trained models is trained at a given user computing device by:
comparing different portions of an aggregated depiction of a stream of non-obfuscated user data gathered at the given user computing device;
adjusting the stream of non-obfuscated user data based on one or more po

Assignees

• Paypal Inc

Inventors

• Chhibber Abhishek
• Desai Darshankumar Bhadrasinh
• Todasco Michael Charles
• Naware Vidyut Mukund
• Sharma Nitin S

Classifications

• G06N3/098

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

• H04L67/01

  Protocols · CPC title

• G06F16/2379

  Updates performed during online database operations; commit processing · CPC title

• G06N5/045

  Explanation of inference; Explainable artificial intelligence [XAI]; Interpretable artificial intelligence · CPC title

• G06Q10/0635

  Risk analysis of enterprise or organisation activities · CPC title