ML Model Fraud Detection via Output Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning models face difficulties in detecting unauthorized use or falsification, which can lead to serious accidents, especially in critical systems, due to the complexity of nonlinear models and large parameter sets, making it challenging to verify internal behavior and recognize minor falsifications.
Innovation Solution
A machine-learning model fraud detection system that includes a license/model management apparatus to manage license, model, and user information, generating a test data-trained model by learning from pre-trained models using test data, and comparing output data from user models with stored values to determine fraudulence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning models with large parameter sets and nonlinear structures are used to improve task execution capability, then model performance is improved, but difficulty in detecting and measuring internal behavior increases
Solution Approach 1:
The patent applies preliminary action by generating test data and establishing expected output values before the model is deployed to users. The fraud detection system pre-prepares verification data and stores expected output values in advance, enabling subsequent fraud detection without requiring internal model inspection. This resolves the contradiction by establishing detection capabilities before the complexity issue arises in deployed models.
2Adaptability or versatility
If machine learning models are falsified by malicious third parties to improve unauthorized use capability, then model adaptability is improved, but reliability deteriorates
Solution Approach 1:
The patent implements feedback by comparing the actual output of the deployed model with pre-stored expected output values. When a user inputs test data, the system receives the model's output, compares it against the expected value, and determines whether fraud has occurred. This closed-loop feedback mechanism detects falsified models while allowing legitimate model adaptations, resolving the contradiction between adaptability and reliability.
3Reliability
If conventional fraud detection methods are used to improve security, then detection capability is improved, but device complexity increases
Solution Approach 1:
The patent extracts the fraud detection function from complex internal model analysis and separates it into an independent verification system. Instead of analyzing the complex nonlinear model internals, the system extracts only the input-output behavior for verification by comparing model outputs with pre-stored expected values. This extraction approach maintains high detection capability while significantly reducing system complexity.
Data Source
AI summary
A machine learning model fraud detection system and fraud detection method wherein a license/model management apparatus: generates a test data-trained model by inputting a pre-trained model and test data associated therewith from a licensor apparatus, carrying out learning using the test data on the pre-trained model; stores the test data-trained model in association with the output values obtained when the test data is executed in the test data-trained model; inputs the associated test data into a user model, executes the model when the user model is inputted from a user apparatus using the test data-trained model; compares the output data from the user model with the stored output values from the test data-trained model and detects the fraud if the resulting error is outside tolerance limits.