VQ-VAE Adversarial Defense for Vehicle Sensor Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current adversarial defense methods for autonomous vehicles, such as Gradient Masking, pre-processing techniques, and Adversarial Training, face limitations including susceptibility to attacks, computational expense, and poor generalization to different adversarial scenarios, making them ineffective for reliable real-time environment perception.

Innovation Solution

A method utilizing a Vector Quantized Variational AutoEncoder (VQ-VAE) for adversarial defense, which maps high-dimensional sensor data into a low-dimensional latent space and applies vector quantization to remove adversarial perturbations, allowing for robust defense against various types of attacks while avoiding posterior collapse issues.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If Gradient Masking is used for adversarial defense, then the network gradients are hidden to prevent attacks, but the defense fails against surrogate model attacks

Engineering Contradiction:
Improveadversarial defense reliabilityVSAvoiddefense effectiveness against different attack types
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent introduces an intermediary defense mechanism that processes sensor data before it reaches the neural network. This intermediary layer acts as a mediator that detects and neutralizes adversarial perturbations without exposing the network's internal gradients, thereby preventing both direct and surrogate model attacks while maintaining defense reliability across different attack types.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent applies preliminary processing to the sensor data to detect and remove adversarial perturbations before the data is fed into the neural network for environment perception. This preliminary defense action prevents adversarial examples from reaching the network, addressing the vulnerability of Gradient Masking methods to surrogate model attacks.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If pre-processing methods are applied to remove adversarial perturbations, then some defense improvement is achieved, but clean images are significantly altered and generalization to different attack types is poor

Engineering Contradiction:
Improveadversarial defense capabilityVSAvoidimage quality preservation
Core Design Contradiction:
ReliabilityVSEase of operation

Solution Approach 1:

The patent applies local quality processing by selectively modifying only the adversarial perturbation components in the sensor data while preserving the clean image content. The defense mechanism identifies and removes malicious perturbations locally without applying blanket transformations that would degrade overall image quality, thus maintaining both defense reliability and image fidelity.

Inventive Principle:
Principle #3Local quality

3Reliability

If Adversarial Training is used to improve robustness, then the neural network learns smoother decision boundaries, but the training becomes computationally very expensive

Engineering Contradiction:
Improvedecision boundary robustnessVSAvoidtraining computational cost
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent extracts and removes adversarial perturbations from the training data before feeding it to the neural network. By taking out the harmful adversarial components and using only clean data for training, the method achieves robust decision boundaries without the computational expense of generating and training on adversarial examples, significantly reducing training energy consumption.

Inventive Principle:
Principle #2Taking out (Extraction)

4Measurement precision

If high-dimensional sensor data is processed directly by neural networks, then detailed environment perception is achieved, but the system becomes vulnerable to adversarial attacks

Engineering Contradiction:
Improveenvironment perception accuracyVSAvoidadversarial attack susceptibility
Core Design Contradiction:
Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The patent converts the vulnerability to adversarial attacks into a benefit by using the defense mechanism to detect and remove perturbations. The high-dimensional sensor data processing is protected by transforming the harmful adversarial input into a detectable anomaly that can be filtered out, thereby maintaining measurement precision while eliminating attack susceptibility.

Inventive Principle:
Principle #22Blessing in disguise (Convert harm into benefit)

Data Source

PatentEP4120136A1Method for automatically executing a vehicle function, method for training a machine learning defense model and defense unit for a vehicle
Publication Date: 2023.01.18 VOLKSWAGEN AG
  • EP4120136A1 patent drawingFigure 1~2
  • EP4120136A1 patent drawingFigure 3
  • EP4120136A1 patent drawingFigure 4

AI summary

Method for automatically executing a vehicle function of a, in particular at least partially autonomously controlled, vehicle (1) based on spatially resolved raw sensor data (I) for environment perception generated by at least one sensor device (22, 24, 26) for environment detection of the vehicle (1), including: - receiving spatially resolved raw sensor data generated by the at least one sensor device (22, 24, 26) of the vehicle (1); - processing sensor data which are characteristic for the spatially resolved raw sensor data (I) by a defense unit (30) for adversarial defense against and/or for detection of adversarial attacks on spatially resolved raw sensor data, the defense unit (30) determining defense output sensor data, wherein for the determination of the defense output sensor data, the sensor data are mapped from a high dimensional input space representation into a low dimensional latent space representation of the sensor data and a vector quantization operation is applied to the latent space representation of the sensor data; - executing the vehicle function based on the defense output sensor data.