Probabilistic Encoder Neural Network Training for Adversarial Robustness

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning models, particularly deep neural networks, are vulnerable to adversarial attacks, which are difficult to defend against due to their non-linear and non-convex nature, and current defense mechanisms are not adaptive, leaving vulnerabilities open to different types of attacks.

Innovation Solution

A neural network training method that incorporates a probabilistic encoder to encode input data into a distribution over a latent space representation, using a multi-objective loss function to maximize shared information between clean and adversarial samples while minimizing view-specific information, thereby enhancing adversarial robustness.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional defense mechanisms like weight decay and dropout are used, then the model structure remains simple, but they do not provide practical defense against adversarial examples

Engineering Contradiction:
Improveadversarial robustnessVSAvoidadaptability to different attack types
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent implements dynamic adaptability by training the neural network to recognize and adapt to multiple types of adversarial attacks (white-box, black-box, transfer attacks) during training. The system dynamically adjusts its defense strategy based on the detected attack type, transforming a static defense mechanism into a dynamic one that can respond to different threat scenarios.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes key parameters of the neural network including introducing an adaptability parameter that controls the model's response to adversarial inputs. By modifying training parameters and using techniques like label smoothing and temperature scaling, the system achieves improved robustness across multiple attack types without requiring complete retraining for each scenario.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If adversarial training is used to generate many adversarial examples for training, then adversarial robustness improves, but computational overhead and training time increase significantly

Engineering Contradiction:
Improveadversarial robustnessVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-training the neural network on a diverse set of adversarial examples covering multiple attack types before deployment. This preliminary training phase equips the model with generalizable defense capabilities, reducing the need for extensive computational resources during actual adversarial encounters and minimizing real-time training requirements.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses partial action by focusing training efforts on the most critical attack vectors and using techniques like selective adversarial training. Instead of exhaustively training against all possible adversarial variations, the system identifies and defends against the most significant threats, achieving robustness with reduced computational overhead.

Inventive Principle:
Principle #16Partial or excessive action

3Adaptability or versatility

If the neural network is trained to defend against multiple types of adversarial attacks, then adaptability improves, but the complexity of the training process and model architecture increases

Engineering Contradiction:
Improveadaptability to different attack typesVSAvoidtraining process complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements universality by designing a single neural network architecture that can defend against multiple types of adversarial attacks (white-box, black-box, transfer attacks) simultaneously. The model uses universal defense mechanisms such as input preprocessing, robust loss functions, and post-processing techniques that apply across different attack scenarios, eliminating the need for separate specialized models for each attack type.

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

Solution Approach 2:

The patent applies segmentation by dividing the defense mechanism into distinct modular components: input validation layer, core neural network, adversarial training module, and output verification layer. Each component handles specific aspects of adversarial defense, allowing independent optimization and simplifying the overall training process while maintaining comprehensive protection against multiple attack types.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4288912B1Method and system for training a neural network for improving adversarial robustness
Publication Date: 2025.08.13 MITSUBISHI ELECTRIC CORP
  • EP4288912B1 patent drawingFigure 1
  • EP4288912B1 patent drawingFigure 2A
  • EP4288912B1 patent drawingFigure 2B(a)~2B(d)

AI summary

Embodiments of the present disclosure disclose a method and a system for training a neural network for improving adversarial robustness. The method includes collecting a plurality of data samples comprising clean data samples and adversarial data samples. The training of the neural network includes training of a probabilistic encoder to encode the plurality of data samples into a probabilistic distribution over a latent space representation. In addition, the training of the neural network comprising training of a classifier to classify an instance of the latent space representation to produce a classification result. In addition, the method includes training shared parameters of a first instance of the neural network using the clean data samples and a second instance of the neural network using the adversarial data samples. Further, the method includes outputting the shared parameters of the first instance of the neural network and the second instance of the neural network.