Trained Model Verification via Rule Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for verifying trained models are limited to algorithms based on decision trees and cannot effectively verify models using other algorithms, such as convolutional neural networks (CNNs), which require a different framework for validation.
Innovation Solution
A trained model verification system that converts models into a rule model with an input/output relationship equivalent to the original, using a verification dataset to verify the model's correctness, allowing for the verification of models based on algorithms other than decision trees.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If verification techniques are limited to decision tree algorithms, then verification can be performed using existing decision tree frameworks, but the system cannot verify mathematical models based on other algorithms such as convolutional neural networks
Solution Approach 1:
The patent introduces a rule model as an intermediary representation that bridges different algorithm types. The rule model convertor transforms various mathematical models (decision trees, CNNs, etc.) into a unified rule model format with if-then rules, which can then be verified using existing decision tree verification frameworks. This intermediary representation enables verification of diverse algorithms without requiring separate verification systems for each algorithm type.
Solution Approach 2:
The patent changes the representation parameter of mathematical models from algorithm-specific formats to a unified rule-based format. By converting models into if-then rules with explicit conditions and outcomes, the system transforms complex algorithmic behaviors into a standardized parameter structure that can be systematically verified, thereby expanding verification capability across different algorithm types.
2Adaptability or versatility
If a unified verification framework is created for all algorithm types, then verification capability is expanded to various mathematical models, but the complexity of the verification system increases due to the need for model conversion and additional processing
Solution Approach 1:
The patent segments the verification system into distinct functional modules: a rule model convertor that handles model transformation, a verification data set generator that creates test data, and a verificator that performs actual verification. This segmentation allows each module to handle specific tasks independently, making the overall complex system more manageable and maintainable while enabling verification of diverse algorithms.
3Adaptability or versatility
If model conversion to rule format is performed, then models beyond decision trees can be verified, but the conversion process requires additional processing steps and computational resources
Solution Approach 1:
The patent performs model conversion to rule format as a preliminary action before verification. By transforming the mathematical model into a rule model in advance, the system prepares the model for verification, allowing the actual verification process to proceed more efficiently. This preliminary conversion step, while requiring additional time, enables the use of optimized verification algorithms and avoids repeated conversion during verification.
Data Source
AI summary
A trained model verification system according to the present disclosed technique includes: a rule model convertor that converts a trained model to be verified into a rule model having an input/output relationship equivalent to that of the trained model; a verification data set generator that generates a verification data set for the rule model; and a verificator that verifies the rule model or the trained model using the verification data set.