Neural Network Adversarial Training Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods to enhance neural network robustness against adversarial attacks, such as adversarial learning, often degrade performance on non-adversarial images.

Innovation Solution

A method that trains neural networks using a learning base comprising both adversarial and non-adversarial images, where adversarial images are paired with target data and non-adversarial images are used to maintain performance on non-adversarial data, allowing for improved robustness without degrading performance on non-adversarial images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If adversarial learning is used to improve robustness against adversarial attacks, then the neural network's robustness is improved, but the performance on non-adversarial images degrades

Engineering Contradiction:
Improverobustness against adversarial attacksVSAvoidperformance on non-adversarial images
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The training process is segmented into two distinct phases: adversarial learning steps where the network learns from adversarial images with target data indicating the intended misclassification, and non-adversarial learning steps where the network learns from normal images. This segmentation allows the network to develop specialized responses for adversarial inputs while maintaining normal performance on legitimate inputs.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The learning method employs periodic alternation between adversarial learning steps and non-adversarial learning steps. During adversarial steps, the network is trained on adversarial images to improve robustness; during non-adversarial steps, it trains on normal images to maintain performance. This periodic switching prevents performance degradation on non-adversarial images while still improving robustness.

Inventive Principle:
Principle #19Periodic action

2Reliability

If the learning base includes only adversarial images, then robustness training is maximized, but the neural network cannot maintain performance on non-adversarial images

Engineering Contradiction:
Improverobustness trainingVSAvoidperformance on non-adversarial images
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The learning base is designed to be universal, serving multiple functions: it includes both adversarial images for robustness training and non-adversarial images for maintaining normal performance. This multi-functional learning base allows the neural network to adapt to different types of inputs and maintain versatility across both adversarial and non-adversarial scenarios.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3920105B1Learning method for a neural network and computer program implementing such a method
Publication Date: 2024.11.20 BULL SA
  • EP3920105B1 patent drawingFigure 1~2
  • EP3920105B1 patent drawingFigure 3

AI summary

The invention relates to a method (300) of supervised adversarial learning of a neural network (102), comprising at least one iteration of a training step (304), called adversarial, comprising the following operations: - provisioning, to said neural network (102) of an image (104), called adversarial, comprising a modification, called adversarial attack, intended to direct said neural network (102) towards a result, called target, different from an expected result, and - provisioning, to said neural network (102), of a first piece of data, called result data, indicating the expected result for said adversarial image (104); characterized in that said adversarial learning step (304) further comprises a provision, to said neural network (102), of a second piece of data, called target data, indicating to said neural network (102) said target.