Neural Network Robustness Verification via Interval Bound Propagation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

The integration of pretrained artificial neural networks in autonomous systems often lacks access to complete training and verification processes, necessitating a method to ensure robustness verification and training for system integrators.

Innovation Solution

A method is provided to train and verify the robustness of artificial neural networks by determining input and output variable limits within predefined intervals, using disturbance models to ensure the network's output remains within permissible bounds, even under disturbances.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a subcontractor performs training and verification of artificial neural networks, then the robustness and reliability of autonomous systems are improved, but system integrators lack access to complete training and verification processes

Engineering Contradiction:

Improverobustness of artificial neural networkVSAvoidaccess to training process information

Core Design Contradiction:

ReliabilityVSLoss of information

Solution Approach 1:

The patent introduces an intermediary verification method that uses interval bound propagation to compute output bounds without requiring access to internal training data. This mediator approach allows verification of robustness through mathematical bounds computation, bridging the information gap between subcontractors who train models and integrators who deploy them.

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent replaces the need for direct access to training processes with a mathematical verification mechanism. Instead of mechanically accessing training data and processes, the system uses interval arithmetic and bound propagation to verify robustness, substituting physical/information access with computational verification.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If interval bound propagation is used for verifiable robustness training, then measurement precision of output bounds is improved, but computational complexity increases

Engineering Contradiction:

Improveprecision of output variable boundsVSAvoidcomputational complexity of verification process

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the verification process into discrete interval bound propagation steps for each layer and dimension of the neural network. By dividing the complex verification task into manageable interval computations for individual layers, the system achieves precise output bounds while making the computational process more tractable and verifiable.

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS12468942B2Method for training and/or verifying a robustness of an artificial neural network

Publication Date: 2025.11.11 ROBERT BOSCH GMBH

AI summary

A device, a method and a computer program for training and/or verifying the robustness of an artificial neural network. The artificial neural network is designed to determine an output variable. The method includes: predefining an input variable for the network which has a plurality of dimensions. For each dimension of the input variable or for each dimension of an output of a linear layer of the artificial neural network without an activation function to which the input variable is mapped by the artificial neural network, the method includes a determination of an upper input variable limit for which a disturbance variable model by which the input variable is able to be mapped to a disturbed input variable has the highest possible value in the dimension, and a determination of a lower input variable limit for which the disturbance variable model has the lowest value possible in the dimension.