Vehicle AI Testing Circuit for Robustness Checks Under Perturbations

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

AI-based information processing systems, particularly deep neural networks, face challenges in transparency, susceptibility to adversarial perturbations, and robustness testing, making it difficult to develop and validate them for autonomous vehicle control, especially in varying environmental conditions.

Innovation Solution

A method and device for checking AI-based information processing systems using a testing circuit that evaluates sensor data and generates test results stored in a multi-dimensional data structure, allowing for continuous assessment and updating of system robustness, including unmonitored testing methods that observe the system without interfering with its operation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If deep neural networks are used for AI-based information processing, then the system can process sensor data and generate control outputs, but the system becomes non-transparent and susceptible to adversarial perturbations

Engineering Contradiction:
Improvesensor data processing capabilityVSAvoidsystem transparency and robustness
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent introduces a testing circuit as an intermediary component that sits between the sensor data input and the AI-based information processing system. This testing circuit applies adversarial perturbations to the sensor data before it reaches the neural network, enabling systematic robustness testing without interfering with the normal operation of the AI system. The testing circuit mediates between the need for robustness verification and the operational integrity of the deep neural network.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Ease of manufacture

If neural networks are trained on synthetic data, then development costs are reduced, but the network performance on real sensor data becomes weak

Engineering Contradiction:
Improvetraining data generation costVSAvoidnetwork performance on real data
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The patent applies preliminary action by performing robustness testing with adversarial perturbations during the training phase or before deployment. The testing circuit pre-evaluates the neural network's vulnerability to various perturbations using synthetic test data, allowing developers to identify and address robustness issues before the network is deployed to process real sensor data. This preliminary robustness verification bridges the gap between synthetic training data and real-world performance.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If extensive robustness testing is performed on neural networks, then system reliability improves, but the complexity of testing and evaluation increases

Engineering Contradiction:
Improverobustness against perturbationsVSAvoidtesting system complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the robustness testing process into distinct functional modules within the testing circuit. The testing circuit is divided into separate components that independently generate different types of adversarial perturbations (e.g., sensor noise, weather influences, image manipulations). Each module can be independently configured and tested, allowing comprehensive robustness evaluation while maintaining manageable system complexity through modular architecture.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentUS11912289B2Method and device for checking an AI-based information processing system used in the partially automated or fully automated control of a vehicle
Publication Date: 2024.02.27 VOLKSWAGEN AG
  • US11912289B2 patent drawing
  • US11912289B2 patent drawing

AI summary

The invention relates to a method for checking an AI-based information processing system used in the partially automated or fully automated control of a vehicle, wherein at least one sensor of the vehicle provides sensor data, the captured sensor data are evaluated by an AI-based information processing system arranged in a first control circuit of the vehicle and, from the evaluated sensor data, at least one output for controlling the vehicle is generated. The AI-based information processing system is checked by a testing circuit arranged in a second control circuit of the vehicle using at least one testing method, and wherein a test result of the at least one testing method is stored, with a reference to the tested AI-based information processing system and to the at least one testing method used, in a multi-dimensional data structure in a database arranged in the vehicle.