Patent · US Active

Training method for robust neural network based on feature matching

US11682194B2 · kind B2 · utility

0Cited by
3References
5Claims
0Family size

Assignee

Inventors

Key dates

Filing dateSep 22, 2022
Grant dateJun 20, 2023
Priority date
Expiry dateSep 22, 2042

Classification

  • Technology area (CPC G)Physics
  • CPC primaryG06V20/70
  • WIPO fieldComputer technology
  • WIPO sectorElectrical engineering

Abstract

A training method for a robust neural network based on feature matching is provided in this disclosure, which includes following steps. Step A, a first stage model is initialized. The first stage model includes a backbone network, a feature matching module and a fullple loss function. Step B, the first stage model is trained by using original training data to obtain a second stage model. Step C, the second stage model is attacked so as to generate PGD adversarial samples of the original training data, and the second stage model is trained again with the generated adversarial samples and the original training data. Step D, training parameters are adjusted and the second stage model is trained again, and parameters for which the model has highest accuracy on an original test set are saved.

Source: USPTO / EPO open patent data. Objective bibliographic and citation counts.