Validation Dataset Masking for AI Cheating Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI models face challenges in ensuring fairness and protecting personally identifiable information (PII) and protected attributes (PA) due to difficulties in collecting robust datasets and securing data, leading to performance issues and the risk of cheating during validation.
Innovation Solution
A system generates a combined dataset with genuine validation data mixed with synthetic data, using statistical techniques and random label permutations to match distributional characteristics, and includes a cheating detection mechanism to identify and prevent model overfitting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If validation data is used to train or retrain the AI model, then performance metrics appear high when using the validation data as input, but the model cannot generalize to new information and performs poorly on real-world data
Solution Approach 1:
The patent creates synthetic copies of validation data that are statistically indistinguishable from real data but contain embedded cheating detection patterns. These synthetic copies are mixed with real validation data to form a combined dataset that appears identical to models, preventing models from detecting and exploiting the validation data while maintaining legitimate performance measurement capabilities.
Solution Approach 2:
The patent introduces an intermediary layer of synthetic data with embedded patterns that acts as a mediator between the validation data and the model. This intermediary prevents direct exploitation of validation data while still allowing legitimate performance evaluation, as the embedded patterns can detect whether the model has accessed the validation data.
2Reliability
If homomorphic encryption is used to protect PII and PA, then data protection is achieved, but time consuming changes to models are required which are not widely deployed
Solution Approach 1:
Instead of modifying models to use homomorphic encryption, the patent creates synthetic copies of the protected data that embed detection patterns. This approach protects PII and PA by replacing them with synthetic alternatives that maintain statistical properties while enabling cheating detection, avoiding the need for complex encryption implementations.
Solution Approach 2:
The patent changes the parameters of the validation data by embedding cheating detection patterns in synthetic copies. This transformation allows the data to maintain its protective function while adding detection capabilities, avoiding the need to implement homomorphic encryption in the models themselves.
3Reliability
If a large amount of synthetic data is generated to protect validation data, then the dataset becomes computationally difficult to exploit, but the complexity of data generation and management increases
Solution Approach 1:
The patent performs preliminary action by embedding cheating detection patterns in synthetic data during the data generation phase. This pre-embedding of patterns simplifies subsequent security verification, as the detection capability is already built into the data structure rather than requiring complex analysis during model validation.
Solution Approach 2:
The patent generates synthetic copies of validation data with embedded patterns, creating a protective layer that is computationally difficult to exploit. These copies maintain the statistical properties of real data while incorporating detection mechanisms, balancing security enhancement with manageable complexity.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Systems and methods for protecting a data in a validation dataset. The system may identify characteristics of a dataset using, for example, a trainable model and may generate fake data based on the identified characteristics of the dataset. The fake data may be interleaved with the validation dataset and may be transmitted to a client for validating against a trained model on a client. A portion of the output from the trained model of the client that corresponds to the validation dataset is may be identified. Metrics may be generated based on the identified portion of the output.