Certification-based robust training by refining decision boundary

Patent metadata

Field	Value
Publication number	US-12567232-B2
Application number	US-202217933473-A
Country	US
Kind code	B2
Filing date	Sep 19, 2022
Priority date	Sep 19, 2022
Publication date	Mar 3, 2026
Grant date	Mar 3, 2026

Abstract

A computer implemented method for certifying robustness of image classification in a neural network is provided. The method includes initializing a neural network model. The neural network model includes a problem space and a decision boundary. A processor receives a data set of images, image labels, and a perturbation schedule. Images are drawn from the data set in the problem space. A distance from the decision boundary is determined for the images in the problem space. A re-weighting value is applied to the images. A modified perturbation magnitude is applied to the images. A total loss function for the images in the problem space is determined using the re-weighting value. A confidence level of the classification of the images in the data set is evaluated for certifiable robustness.

First claim

What is claimed is:

1 . A computer implemented method for certifying robustness of image classification in a neural network, comprising:
initializing a neural network model in the neural network, wherein the neural network model includes a problem space and a decision boundary in the problem space; receiving, by a processor, a data set of images and image labels; applying a user input perturbation schedule to the data set of images and image labels; drawing images from the data set in the problem space using the user input perturbation schedule; determining a distance from the decision boundary for the images in the problem space; applying a re-weighting value to the images in the problem space, wherein the re-weighting value affects the distance to the decision boundary for at least some of the images in the problem space; determining a total loss function for the images in the problem space using the re-weighting value; applying a modified perturbation magnitude to the images in the problem space; determining positions of the images in the problem space relative to the decision boundary, using the modified perturbation magnitude; evaluating a confidence level of a classification of the images in the data set based on the total loss function and the positions of the images in the problem space relative to the decision boundary, using the modified perturbation magnitude; and terminating training of the neural network model.

2 . The method of claim 1 , further comprising using a symmetrical re-weighting function to apply the re-weighting value to the images.

3 . The method of claim 2 , further comprising identifying worst case data points proximate the decision boundary, wherein the symmetrical weighting function is based on the distance from the decision boundary for the worst case data points.

4 . The method of claim 1 , further comprising customizing the modified perturbation magnitude to individual images in the problems space.

5 . The method of claim 1 , further comprising automatically tuning the modified perturbation magnitude to individual images in the problems space.

6 . The method of claim 1 , further comprising customizing the modified perturbation magnitude for individual images based on a confidence level of the distance of individual images to the decision boundary.

7 . The method of claim 1 , further comprising:
determining a gradient descent direction from the total loss function; and updating neural network parameters using the gradient descent direction.

8 . A non-transitory computer program product for certifying robustness of image classification in a neural network, the computer program product comprising:
one or more computer readable storage media, and program instructions collectively stored on the one or more computer readable storage media, the program instructions comprising: initializing a neural network model in the neural network, wherein the neural network model includes a problem space and a decision boundary in the problem space; receiving, by a processor, a data set of images and image labels; applying a user input perturbation schedule to the data set of images and image labels; drawing images from the data set in the problem space using the user input perturbation schedule; determining a distance from the decision boundary for the images in the problem space; applying a re-weighting value to the images in the problem space, wherein the re-weighting value affects the distance to the decision boundary for at least some of the images in the problem space; determining a total loss function for the images in the problem space using the re-weighting value; applying a modified perturbation magnitude to the images in the problem space; determining positions of the images in the problem space relative to the decision boundary, using the modified perturbation magnitude; evaluating a confidence level of a classification of the images in the data set based on the total loss function and the positions of the images in the problem space relative to the decision boundary, using the modified perturbation magnitude; and terminating training of the neural network model.

9 . The computer program product of claim 8 , wherein the program instructions further comprise using a symmetrical re-weighting function to apply the re-weighting value to the images.

10 . The computer program product of claim 9 , wherein the program instructions further comprise identifying worst case data points proximate the decision boundary, wherein the symmetrical weighting function is based on the distance from the decision boundary for the worst case data points.

11 . The computer program product of claim 8 , wherein the program instructions further comprise customizing the modified perturbation magnitude to individual images in the problems space.

12 . The computer program product of claim 8 , wherein the program instructions further comprise automatically tuning the modified perturbation magnitude to individual images in the problems space.

13 . The computer program product of claim 8 , wherein the program instructions further comprise customizing the modified perturbation magnitude for individual images based on a confidence level of the distance of individual images to the decision boundary.

14 . The computer program product of claim 8 , wherein the program instructions further comprise:
determining a gradient descent direction from the total loss function; and updating neural network parameters using the gradient descent direction.

15 . A computer server configured to certify robustness of image classification in a neural network, comprising:
a network connection; one or more computer readable storage media; a processor coupled to the network connection and coupled to the one or more computer readable storage media; and a computer program product comprising program instructions collectively stored on the one or more computer readable storage media, the program instructions comprising: initializing a neural network model in the neural network, wherein the neural network model includes a problem space and a decision boundary in the problem space; receiving, by a processor, a data set of images and image labels; applying a user input perturbation schedule to the data set of images and image labels; drawing images from the data set in the problem space using the user input perturbation schedule; determining a distance from the decision boundary for the images in the problem space; applying a re-weighting value to the images in the problem space, wherein the re-weighting value affects the distance to the decision boundary for at least some of the images in the problem space; determining a total loss function for the images in the problem space using the re-weighting value; applying a modified perturbation magnitude to the images in the problem space; determining positions of the images in the problem space relative to the decision boundary, using the modified perturbation magnitude; evaluating a confidence level of a classification of the images in the data set based on the total loss function and the positions of the images in the problem space relative to the decision boundary, using the modified perturbation magnitude; and terminating training of the neural network model.

16 . The computer server of claim 15 , wherein the program instructions further comprise using a symmetrical re-weighting function to apply the re-weighting value to the images.

17 . The computer server of claim 16 , wherein the

Assignees

• IBM
• Massachusetts Inst Technology

Inventors

• Nguyen Lam Minh
• Zhang Wang
• Das Subhro
• Chen Pin-Yu
• Megretski Alexandre
• Daniel Luca

Classifications

• G06V10/774

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V10/764 Primary

  using classification, e.g. of video objects · CPC title

• G06V10/82 Primary

  using neural networks · CPC title