ML Model Assurance Scoring Through Automated Adversarial Testing

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing ML models are susceptible to adversarial attacks, leading to inaccurate results and security breaches, particularly in autonomous systems where human oversight is limited or absent.

Innovation Solution

An automated system generates tailored security tests based on the type and format of ML models to assess their robustness and resilience, providing a model assurance score (MAS) that guides potential modifications to enhance security.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If automated security testing is implemented for ML models, then the ability to detect adversarial attacks is improved, but the complexity of the testing system increases

Engineering Contradiction:

Improveability to withstand adversarial attacksVSAvoidcomplexity of security testing system

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The security testing system is segmented into distinct modules: model analysis module that identifies ML model type and characteristics, test case generation module that creates tailored security tests, and execution module that runs the tests. This segmentation allows each component to specialize in specific tasks, improving detection capability while managing system complexity through modular design.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system performs preliminary analysis of the ML model to identify its type, framework, and architecture characteristics before generating security test cases. This preliminary action ensures that the subsequent security testing is tailored to the specific model properties, improving effectiveness without requiring a universally complex testing framework.

Inventive Principle:

Principle #10Preliminary action

2Measurement precision

If tailored security tests are generated for specific ML model types, then the precision of security assessment is improved, but the time required for testing increases

Engineering Contradiction:

Improveprecision of model assurance scoreVSAvoidtime for security testing

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system dynamically adjusts testing parameters based on the identified ML model type and characteristics. Different model types (e.g., neural networks, decision trees, support vector machines) receive customized test case sets with appropriate parameters, achieving high precision assessment while optimizing testing duration for each specific model category.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS20260003971A1Automated machine learning model assessment

Publication Date: 2026.01.01 SAUDI ARABIAN OIL CO

AI summary

The present disclosure relates to computer-implemented methods, software, and systems for security testing of machine learning (ML) models. A request is received to perform a security test on a first ML model, wherein the request comprises a file with the first ML model. The first ML model is analyzed to identify a type of the first ML model. Applicable test cases are generated for testing the first ML model. The applicable test cases are executed to determine a model assurance score indicative of an ability of the first ML model to withstand adversarial attacks. A report is provided for display at a display device, the report comprising the model assurance score for the first ML model.