Synthetic Entity Replacement for Privacy-Preserving Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating machine-learning model training data from redacted information fail to maintain relationships between entities and often introduce unintended bias, leading to inaccurate training and potential exposure of sensitive personal information.
Innovation Solution
A system that generates replacement entities for redacted information while maintaining relationships and controlling bias by using frequency distributions and structured dataset analysis to ensure accurate and privacy-preserving training data generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If redacted information is used to generate training data, then privacy protection is improved, but relationship accuracy between entities deteriorates
Solution Approach 1:
The patent creates synthetic training data by copying the structural relationships from redacted information documents while replacing sensitive personal identifiers with anonymized placeholders. This allows the training data to maintain the relational patterns and contextual information needed for accurate machine learning while removing identifiable personal information, thus resolving the contradiction between privacy protection and relationship accuracy.
Solution Approach 2:
The patent extracts and separates the relational structure from the personal identification information in the documents. By taking out the sensitive PII (personal identifiable information) while preserving the entity relationships and contextual data, the system generates training data that maintains accuracy for relationship modeling while protecting privacy.
2Object-affected harmful factors
If redacted information is used to generate training data, then privacy protection is improved, but training accuracy deteriorates
Solution Approach 1:
The patent copies the essential structural and contextual information from redacted documents while replacing sensitive content with anonymized placeholders. This preserves the patterns, relationships, and training-relevant features needed for accurate model training while eliminating PII, thus maintaining training accuracy under privacy constraints.
Solution Approach 2:
The patent changes the parameter of information representation by transforming real personal identifiers into anonymized placeholders while preserving the relational and contextual parameters. This parameter transformation allows the data to remain useful for training accuracy while achieving privacy protection.
3Productivity
If replacement entities are generated without frequency distribution guidance, then generation speed is improved, but bias control deteriorates
Solution Approach 1:
The patent incorporates frequency distribution analysis as a feedback mechanism that guides the replacement entity generation process. By analyzing the frequency of entities in the original data and using this feedback to select appropriate replacements, the system maintains diversity and reduces bias while generating training data at efficient speeds.
Solution Approach 2:
The patent changes the parameter of entity selection by using frequency distribution statistics to guide replacement choices. This parameter-based approach ensures that generated entities reflect the original data's distribution characteristics, controlling for bias while maintaining efficient generation throughput.
Data Source
AI summary
One embodiment provides a computer implemented method, including: obtaining an information document corresponding to an entity, wherein the information document includes redacted information spans; identifying an entity type for each of the redacted information spans, wherein the entity type identifies a relationship between a redacted information span and at least one other entity within the information document; replacing the redacted information spans with replacement entities corresponding to the entity type of a given redacted information span, wherein the replacing is performed in view of a frequency distribution of actual information and wherein the replacing includes maintaining relationships of the redacted information spans; and controlling bias within the replacement entities, wherein the controlling includes detecting bias within the replacement entities.