Deidentified Content Surrogates for Privacy-Safe ML Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical data privacy laws and contractual limitations restrict the long-term storage and usage of user data for training machine learning models, necessitating a method to deidentify personal information while allowing secure sharing for training purposes.
Innovation Solution
A two-step hashing process using a first hashing algorithm (H1) to replace personal information with mathematical representations and a second hashing algorithm (H2) to generate surrogates, enabling secure sharing of deidentified content for training machine learning models without risking reversal to personal information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If personal information is stored and used for training machine learning models, then model training accuracy is improved, but data privacy compliance deteriorates
Solution Approach 1:
The patent extracts personal information from the data and replaces it with hashed mathematical representations. This separation allows the data to be used for model training while the personal information remains securely stored in hash form that cannot be reversed, thus maintaining both training accuracy and privacy compliance.
Solution Approach 2:
The patent introduces hashing algorithms (H1 and H2) as intermediaries between personal information and machine learning models. These hash functions transform personal information into mathematical representations that preserve statistical properties for training purposes while making the original information irretrievable, serving as a mediator that satisfies both training needs and privacy requirements.
2Device complexity
If deidentification is performed using a single hashing algorithm, then processing simplicity is maintained, but security against reversal deteriorates
Solution Approach 1:
The patent segments the hashing process into two distinct algorithms: H1 for initial hashing of personal information and H2 for generating surrogates from the hashed representations. This segmentation creates multiple layers of security where each hash function protects against reversal of the previous one, while maintaining clear separation of concerns in the processing pipeline.
Solution Approach 2:
The patent transitions from a single-dimensional hashing approach to a two-dimensional hashing structure. The first dimension (H1) transforms personal information to mathematical representations, and the second dimension (H2) transforms those representations to surrogates. This dimensional expansion creates additional security layers without significantly complicating the overall process.
Data Source
AI summary
A method, computer program product, and computing system for processing raw content to identify personal information; replacing the personal information within the raw content with a first mathematical representation of the personal information generated using a first hashing algorithm, thus defining deidentified content; and defining a selected surrogate for the first mathematical representation, wherein the selected surrogate has a second mathematical representation generated using a second hashing algorithm that is equivalent to the first mathematical representation of the personal information that was generated using the first hashing algorithm.