Supercharge Your Innovation With Domain-Expert AI Agents!

Security release-oriented aggregation model training method, training device and system

A model training, security-oriented technology, applied in the field of privacy protection, can solve the problems of insecurity and cannot resist the stealing of eavesdroppers' privacy, and achieve the effect of safe and reliable training process

Pending Publication Date: 2022-04-26
国网智能电网研究院有限公司南京分公司 +4
View PDF0 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0005] In order to solve the technical problem that the existing technology cannot resist the eavesdropper's stealing of privacy during the model training process, and the training process is not safe due to the existence of the stolen text, the present invention provides an aggregation model training oriented to security release methods, including:

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Security release-oriented aggregation model training method, training device and system
  • Security release-oriented aggregation model training method, training device and system
  • Security release-oriented aggregation model training method, training device and system

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0099] In order to solve the deficiencies of the prior art, the present invention provides a security release-oriented aggregation model training method implemented by the server, such as figure 1 shown, including:

[0100] Step 1. The server publishes the processed deep learning model to the client that provides training data;

[0101] Step 2. The server aggregates the received gradient information provided by multiple clients using the security aggregation protocol to obtain the model gradient;

[0102] Step 3. The server updates the deep learning model based on the model gradient, and sends the updated deep learning model to the client, and continues iterative training until the model training is completed;

[0103] Wherein, the gradient information is obtained by each client using local data to train the deep learning model based on distributed differential privacy technology.

[0104] Among them, in step 1, the server publishes the processed deep learning model to the c...

Embodiment 2

[0124] The present invention also provides a safety release-oriented aggregation model training method implemented by the client, such as figure 2 shown, including:

[0125] Step 1. The client uses local data to train the deep learning model sent by the server based on distributed differential privacy technology;

[0126] Step 2. The client processes the training results based on the security aggregation protocol to obtain gradient information, and sends the gradient information to the server;

[0127] Step 3, the client continues to iteratively train the updated deep learning model sent by the server until the training of the deep learning model is completed;

[0128] Wherein, the ciphertext sent by the server is obtained by aggregating the model gradients provided by multiple clients by the server; the updated deep learning model is obtained by aggregating the plaintext information provided by the server based on multiple clients The gradient is obtained by updating the d...

Embodiment 3

[0153] The present invention also provides a security release-oriented aggregation model training system, which is composed of a plurality of training devices, and the training devices include a server and a plurality of clients providing training data;

[0154] The server includes: model release module, gradient determination module and model update module.

[0155] Model publishing module, for publishing the processed deep learning model to multiple clients providing training data;

[0156] The gradient determination module is used to aggregate the received gradient information provided by multiple clients using the security aggregation protocol to obtain the model gradient; it is specifically used for:

[0157] The server performs homomorphic addition operation on the received encrypted model gradients provided by multiple clients to obtain the ciphertext, and sends the ciphertext to each client;

[0158] When the server receives the plaintext information returned by each ...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention discloses a security release-oriented aggregation model training method, training device and system, the training device comprises a server and clients, each client obtains federal learning overall information in a training release process, and then uses local data to train a deep learning model by using a distributed differential privacy technology; the server side uses a security aggregation protocol to complete model aggregation, and then transmits new model parameters to the client side for iterative training; according to the method, the finally published model can meet the differential privacy of a model observer, the sub-model gradient in the training process is invisible to a data provider, and a server cannot distinguish a data source. Therefore, on one hand, the privacy stealing attack of a server side single party or other clients participating in training can be resisted, and on the other hand, the method also has certain attack resisting capability for an intermediate eavesdropper or a malicious destructor.

Description

technical field [0001] The present invention relates to the technical field of privacy protection, and in particular to a security release-oriented aggregation model training method, training equipment and system. Background technique [0002] In the context of the rapid development of artificial intelligence technology and the gradual improvement of society and citizens' privacy awareness, federated learning, a deep learning technology, has emerged. One of the most important characteristics of deep learning technology is the need for big data support. A large amount of user data in real scenarios must be put into deep learning training to obtain a satisfactory deep learning model, so how to solve the "data island" problem is very important. The most immediate problem that needs to be faced in collecting data from a large number of scattered users is that users are unwilling to provide data directly to others due to privacy reasons. [0003] In the process of applying a la...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
IPC IPC(8): G06F21/62G06K9/62G06N20/00
CPCG06F21/6245G06F21/6236G06F21/6227G06N20/00G06F18/214
Inventor 石聪聪何维民张舸黄秀丽华景煜
Owner 国网智能电网研究院有限公司南京分公司
Features
  • R&D
  • Intellectual Property
  • Life Sciences
  • Materials
  • Tech Scout
Why Patsnap Eureka
  • Unparalleled Data Quality
  • Higher Quality Content
  • 60% Fewer Hallucinations
Social media
Patsnap Eureka Blog
Learn More

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic, Popular Technical Reports.

© 2025 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap|About US| Contact US: help@patsnap.com