ML Model Fuzz Testing Using Generative Dataset Mutation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing software fuzzing tools are inadequate for testing machine learning models, which are difficult to interpret and debug, and verification is computationally expensive, especially in identifying vulnerabilities and errors in attack scenarios.

Innovation Solution

Implement fuzz testing of machine learning models by generating mutated datasets using generative models to train and retrain models, adjusting parameters based on accuracy, and detecting malicious activity through statistical distribution analysis.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If existing software fuzzing tools are used to test machine learning models, then testing can be performed, but the tools are inadequate for identifying vulnerabilities and errors in attack scenarios

Engineering Contradiction:

Improvevulnerability detection capabilityVSAvoidsuitability for ML model testing

Core Design Contradiction:

ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent transforms traditional software fuzzing parameters into ML-specific parameters by generating mutated training datasets with modified statistical distributions. Instead of fuzzing model inputs directly, the system mutates training data characteristics (distribution, statistics, patterns) to create adversarial training examples that expose model vulnerabilities in attack scenarios.

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent introduces generative models as intermediaries between the fuzzer and the target ML model. The generative model receives mutation seeds and generates mutated training datasets that serve as the actual test input, bridging the gap between traditional fuzzing techniques and ML model testing requirements.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Measurement precision

If machine learning models are trained with more data to improve accuracy, then detection performance improves, but verification becomes computationally expensive

Engineering Contradiction:

Improvedetection accuracyVSAvoidcomputational cost

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent performs preliminary vulnerability detection through fuzzing during the training phase by generating mutated training datasets before final model deployment. This preliminary testing identifies and corrects vulnerabilities early, reducing the need for expensive post-deployment verification and retraining cycles.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system implements feedback loops where fuzzing results from mutated training datasets inform model retraining. When vulnerabilities are detected through statistical distribution analysis of test inputs, the model is retrained with corrected data, creating a continuous improvement cycle that reduces computational costs over time by preventing vulnerability accumulation.

Inventive Principle:

Principle #23Feedback

3Adaptability or versatility

If machine learning models are made more complex to detect sophisticated malicious activity, then detection capability improves, but models become difficult to interpret and debug

Engineering Contradiction:

Improvemalicious activity detection capabilityVSAvoidmodel interpretability

Core Design Contradiction:

Adaptability or versatilityVSDifficulty of detecting and measuring

Solution Approach 1:

The patent implements self-service debugging through automated fuzzing and statistical distribution analysis. The system automatically generates mutated test cases, executes them against the model, detects vulnerabilities through distribution shifts, and triggers retraining without manual intervention, making complex model debugging accessible without deep expertise.

Inventive Principle:

Principle #25Self-service

4Ease of operation

If traditional fuzzing is applied to ML models by mutating input data, then testing can be performed, but it fails to identify vulnerabilities in attack scenarios

Engineering Contradiction:

Improvetesting capabilityVSAvoidvulnerability identification effectiveness

Core Design Contradiction:

Ease of operationVSReliability

Solution Approach 1:

The patent fundamentally changes the fuzzing parameter from input data mutation to training data distribution mutation. Instead of mutating individual input examples, the system mutates the statistical distribution of entire training datasets, creating fundamental shifts in model behavior that expose vulnerabilities in attack scenarios while maintaining ease of operation through automated pipeline execution.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS12500925B2Fuzz testing of machine learning models to detect malicious activity on a computer

Publication Date: 2025.12.16 SOPHOS LTD

AI summary

An example method can include, obtaining information about a machine learning (ML) model configured to detect malicious activity on a computer system, wherein the information includes one or more of a model type of the ML model, an output type of the ML model, or a type of malicious activity that the model is trained to detect, receiving a training dataset, wherein the training dataset includes a plurality of unlabeled examples, generating an additional dataset based on the training dataset using a generative model, wherein the additional dataset includes a plurality of additional unlabeled examples, and, training the machine learning model to generate labels for each example in the training dataset and the additional dataset, using a combination of the training dataset and the additional dataset, wherein the training includes adjusting one or more parameters of the machine learning model based on accuracy of the generated labels.