A multi-source heterogeneous data distillation method and device for data privacy protection

By preprocessing and convolutional neural network model training on multi-source heterogeneous image data sets, combined with unsupervised clustering and two-layer network model optimization technology, the problem that existing technology is difficult to retain the implicit structural characteristics of the data set is solved, and the accuracy of efficient data distillation and classification tasks is improved.

CN118072119BActive Publication Date: 2025-04-08ZHEJIANG LAB
2 Cites 0 Cited by

Patent Information

Application Number
CN202410055534.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-01-15
Publication Date
2025-04-08
Estimated Expiration
2044-01-15

AI Technical Summary

Technical Problem

Existing data distillation methods are difficult to effectively retain the implicit structural features of multi-source heterogeneous image data sets, especially while data compression, it is impossible to effectively distinguish the different structural features of the data set.

Method used

The multi-source heterogeneous data distillation method for data privacy protection is adopted. By acquiring the multi-source heterogeneous image data sets for preprocessing, the convolutional neural network model is trained to extract the characterization data, and the unsupervised clustering and a two-layer network model are optimized to generate the optimized synthetic data set.

Benefits of technology

The implicit structural characteristics of the multi-source data set were successfully retained, and higher accuracy was achieved in downstream classification tasks, effectively achieving the purpose of data distillation.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The present invention discloses a multi-source heterogeneous data distillation method and device for data privacy protection, the method comprising: firstly obtaining a multi-source heterogeneous image data set to be distilled and performing corresponding data preprocessing; secondly using a convolutional neural network to obtain image representation and then clustering the representation data using an unsupervised clustering method; then constructing a double-layer network model that satisfies both distribution matching and parameter matching, matching the original data points corresponding to the centroid points in each cluster as a synthetic data set, iteratively updating the synthetic data set and network parameters through an asynchronous update mechanism until the model converges; finally obtaining the distilled image data set. The present invention can accurately identify the implicit patterns of multi-source heterogeneous image data sets, and synthesize the most representative image data set from each implicit pattern, ultimately achieving the purpose of data privacy protection and data compression.
Need to check novelty before this filing date? Find Prior Art