ML Data Anonymization With Context-Preserving Entity Replacement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning processing by third parties may expose sensitive information, and traditional anonymization techniques can compromise the effectiveness of the model output or make deanonymization difficult.
Innovation Solution
Anonymize data by replacing sensitive entities with contextually meaningful replacement entities, generating a mapping for deanonymization, and using this mapping to reintroduce sensitive entities in model output.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If traditional redaction techniques are used to anonymize sensitive information, then sensitivity protection is improved, but machine learning processing effectiveness deteriorates due to loss of context and semantic meaning
Solution Approach 1:
The patent introduces placeholder tokens as intermediary elements that replace sensitive information during processing. These placeholders serve as mediators between the need for anonymization and the need for contextual preservation, allowing machine learning models to process anonymized data while maintaining structural and contextual relationships. The mapping between placeholders and original sensitive information enables subsequent restoration without exposing sensitive data during processing.
Solution Approach 2:
The patent creates copies of sensitive information in the form of placeholder tokens that preserve the structural and contextual properties of the original data. Instead of completely removing or redacting sensitive information, the system creates representative copies (placeholders) that maintain the data's positional, grammatical, and contextual relationships while eliminating actual sensitive content during processing.
2Object-affected harmful factors
If sensitive information is completely removed or redacted, then sensitivity protection is improved, but the utility and effectiveness of third-party machine learning processing deteriorates
Solution Approach 1:
Placeholder tokens act as intermediaries that allow machine learning models to process data structures and relationships without accessing actual sensitive information. The placeholders maintain the necessary contextual framework for effective processing while the mapping mechanism enables restoration of sensitive information only where needed in the output, preserving both protection and productivity.
Solution Approach 2:
The patent segments the data processing into distinct phases: anonymization phase where sensitive information is replaced with placeholders for processing, and restoration phase where sensitive information is selectively restored using the mapping. This segmentation allows third-party models to process effectively without accessing sensitive data, while the owner retains control over when and how sensitive information is restored.
3Object-affected harmful factors
If anonymization is performed without preserving mapping information, then sensitivity protection is improved, but deanonymization capability deteriorates
Solution Approach 1:
The patent extracts and separates the mapping information (the relationship between placeholders and sensitive information) from the processing workflow. This extracted mapping is stored securely and used only when needed for restoration, allowing the processing itself to occur without sensitive data while preserving the capability to restore sensitive information selectively when required.
Data Source
AI summary
Aspects of the present disclosure relate to a machine learning data anonymizer. To anonymize data that is provided for third party processing, sensitive entities are identified therein, which are replaced with replacement entities accordingly. In examples, the replacement entities include an indication of a category corresponding to the sensitive entity, thereby retaining a context/semantic meaning of the sensitive entity without providing the sensitive entity itself. A mapping is generated that associates replacement entities and corresponding sensitive entities, thereby facilitating subsequent deanonymization. Once generated output is received from the third party (e.g., as may have been generated by a machine learning model), the generative output is processed according to the mapping to substitute replacement entities therein with corresponding sensitive entities, thereby generating deanonymized model output in which sensitive entities are reintroduced and thus available for subsequent processing.