Prediction Model Robustness Bounds Using PSO and IBP

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for quantifying the robustness of deep neural networks in critical applications are computationally inefficient and face scalability issues, especially for large-scale networks, making them unsuitable for real-world deployment in safety-critical tasks.

Innovation Solution

A method using Particle Swarm Optimization (PSO) to estimate a lower bound of the Lipschitz constant for neural networks, combined with Interval Bound Propagation (IBP), enabling efficient quantification of robustness by determining a prediction radius and bounds for perturbed input data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If heuristic search approach with large number of unguided perturbations is used to quantify robustness, then robustness quantification can be performed, but computational efficiency deteriorates due to curse of dimensionality and large computation resources required

Engineering Contradiction:
Improverobustness quantification accuracyVSAvoidcomputational efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent transforms the robustness quantification problem from a search-based approach to a parameter-optimization approach. By formulating the problem as minimizing the maximum output difference over perturbed inputs and using gradient-based optimization with ReLU-specific gradient handling, the method efficiently computes robustness bounds without exhaustive sampling, thus improving computational efficiency while maintaining reliability

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent replaces the mechanical heuristic search process with an automated gradient-based optimization system. By deriving specific gradient computation rules for ReLU networks and implementing automated optimization loops, the system substitutes manual or random perturbation methods with a systematic mathematical optimization approach, significantly reducing computational resources required

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Reliability

If formal verification techniques are used to certify robustness of neural networks, then robustness guarantee is obtained, but device complexity increases and scalability deteriorates for large-sized networks

Engineering Contradiction:
Improverobustness certificationVSAvoidverification system complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies local quality by deriving gradient computation rules specific to ReLU activation functions. By treating ReLU units with specialized gradient handling (zero gradient for negative inputs, one for positive inputs) rather than generic verification methods, the approach simplifies the verification process for networks with this common activation function, reducing complexity while maintaining certification capability

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent changes the verification approach from general-purpose formal methods to a parameterized optimization framework tailored for ReLU networks. By formulating robustness certification as an optimization problem with specific gradient rules for ReLU, the method reduces computational complexity and improves scalability compared to universal formal verification techniques

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If deep and complicated network architecture is used to improve model performance, then prediction accuracy is improved, but finding minimum distortion of adversarial examples becomes more computationally demanding and scalability issues worsen

Engineering Contradiction:
Improveprediction accuracyVSAvoidrobustness evaluation speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent replaces exhaustive adversarial search methods with gradient-based optimization. By computing gradients of the network output with respect to inputs and using these gradients to guide perturbation generation, the system efficiently finds adversarial examples even in deep complicated networks, maintaining evaluation speed while handling complex architectures

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent introduces parameter-specific gradient computation rules for ReLU networks that scale efficiently with network depth and complexity. By deriving closed-form gradient expressions that account for ReLU activation characteristics, the method enables fast robustness evaluation of deep networks without the computational burden of generic verification approaches

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4645170A1Quantifying robustness of prediction models
Publication Date: 2025.11.05 SIEMENS AG
  • EP4645170A1 patent drawingFigure 1
  • EP4645170A1 patent drawingFigure 2
  • EP4645170A1 patent drawing

AI summary

Methods and systems for efficiently quantifying the robustness of a prediction model are disclosed. Embodiments include providing an input data to be predicted, a measure of distortion, and a distortion norm p; performing a forward pass through the prediction model to obtain a prediction value based on the input data; determining, using a Particle Swam Optimization, PSO, algorithm, a maximum value of a q-norm of a gradient of the prediction model within a bounded region around the input data, the bounded region being defined by the measure of distortion and the distortion norm p; calculating a prediction radius of the prediction model based on the determined maximum value and the measure of distortion; and determining a lower and upper bound of a prediction range of the prediction model based on the prediction value and the prediction radius, the lower and upper bounds representing possible values the prediction model is able to output for the given input data.