Neural Network Gradient Adversarial Training for Robust Backpropagation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural networks face challenges in robustness against adversarial attacks, knowledge distillation efficiency, and effective multitask learning, with a lack of formalism in gradient tensor utilization during backpropagation.

Innovation Solution

Gradient adversarial training (GREAT) involves an auxiliary network that classifies gradient tensors during backpropagation, passing an adversarial signal to the main network to regularize weights, enhancing robustness and alignment of gradient tensors.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If standard backpropagation is used for training, then the training process is simple, but the neural network lacks robustness against adversarial attacks

Engineering Contradiction:
Improverobustness against adversarial attacksVSAvoidtraining process complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

An auxiliary neural network is introduced as an intermediary component that processes gradient tensors during backpropagation. This auxiliary network classifies gradient tensors and generates adversarial signals that are fed back to the main network, thereby enhancing robustness without fundamentally altering the core training procedure

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system implements a feedback mechanism where the auxiliary network processes gradient tensors and returns adversarial signals to the main network. This closed-loop feedback allows the network to learn from adversarial examples generated during training, improving robustness while maintaining a relatively simple training framework

Inventive Principle:
Principle #23Feedback

2Productivity

If gradient tensors are used directly during backpropagation, then the process is straightforward, but knowledge distillation efficiency is poor

Engineering Contradiction:
Improveknowledge distillation efficiencyVSAvoidgradient utilization complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The auxiliary neural network serves as a mediator between the student and teacher networks during knowledge distillation. It processes gradient tensors from both networks and facilitates more effective knowledge transfer by aligning gradient distributions, thereby improving distillation efficiency

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system transforms gradient tensors through the auxiliary network, changing their distribution and characteristics. This parameter transformation aligns the gradient distributions between teacher and student networks, enabling more effective knowledge distillation

Inventive Principle:
Principle #35Parameter changes

3Adaptability or versatility

If standard training procedures are used, then the implementation is simple, but multitask learning effectiveness is limited

Engineering Contradiction:
Improvemultitask learning effectivenessVSAvoidtraining procedure complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The auxiliary neural network performs multiple functions: it classifies gradient tensors, generates adversarial signals, and aligns gradient distributions across different tasks. This multi-functional component enables effective multitask learning without requiring separate mechanisms for each task

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

Solution Approach 2:

The system dynamically adjusts gradient tensor characteristics through the auxiliary network, adapting gradient distributions to suit different task requirements. This parameter adaptation enables the network to effectively learn multiple tasks simultaneously

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3794516B1Gradient adversarial training of neural networks
Publication Date: 2025.12.24 MAGIC LEAP INC
  • EP3794516B1 patent drawingFigure 1A
  • EP3794516B1 patent drawingFigure 1B
  • EP3794516B1 patent drawingFigure 1C

AI summary

Systems and methods for gradient adversarial training of a neural network are disclosed. In one aspect of gradient adversarial training, an auxiliary neural network can be trained to classify a gradient tensor that is evaluated during backpropagation in a main neural network that provides a desired task output. The main neural network can serve as an adversary to the auxiliary network in addition to a standard task-based training procedure. The auxiliary neural network can pass an adversarial gradient signal back to the main neural network, which can use this signal to regularize the weight tensors in the main neural network. Gradient adversarial training of the neural network can provide improved gradient tensors in the main network. Gradient adversarial techniques can be used to train multitask networks, knowledge distillation networks, and adversarial defense networks.