Context-Aware Text Sanitization for Interpretable Sensitive Data Masking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional text sanitization techniques, such as regex-based processing, struggle with high false positives, inability to detect slightly changed text, and inability to capture granular data types, resulting in sanitized text that lacks context and is difficult to interpret.
Innovation Solution
A context-aware text sanitization pipeline using named entity recognition and question-answer generation to generate multi-field data structures that mask sensitive information with context-based labels, employing a multiple-choice question reading comprehension model to retain interpretability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional regex-based text sanitization is used, then sensitive information can be masked, but the sanitized text loses context and becomes difficult to interpret
Solution Approach 1:
The patent applies different masking strategies to different portions of text based on their sensitivity and contextual importance. Instead of uniformly masking all sensitive data, the system selectively masks only the necessary portions while preserving contextual information that aids interpretation. This local differentiation resolves the contradiction by maintaining reliability in masking while minimizing information loss.
Solution Approach 2:
The sanitization process is segmented into multiple stages: identification of sensitive information, determination of masking necessity, selective masking application, and context preservation. This segmentation allows the system to handle different types of sensitive data differently, masking critical information while preserving contextual cues, thereby resolving the contradiction between reliable masking and information retention.
2Productivity
If traditional regex-based text sanitization is used, then processing is simple and fast, but false positives are high and granular data types cannot be detected
Solution Approach 1:
The system dynamically adjusts its detection and masking approach based on the specific context and type of sensitive information encountered. Rather than using static regex patterns, the system employs adaptive algorithms that learn from context, allowing it to maintain high processing speeds while improving detection accuracy and reducing false positives through contextual awareness.
Solution Approach 2:
The patent changes the parameters of text analysis from simple pattern matching to contextual parameter evaluation. By considering multiple parameters such as data type, context, sensitivity level, and positional information, the system achieves higher measurement precision in detecting sensitive information while maintaining productivity through efficient parameter-based processing.
3Ease of manufacture
If traditional regex-based text sanitization is used, then implementation is straightforward, but the system cannot detect slightly changed text or non-alphanumeric data types
Solution Approach 1:
The patent implements a universal text sanitization framework that can handle multiple data types and formats through a single system. This multi-functional approach detects and masks various sensitive information types including alphanumeric, non-alphanumeric, and slightly modified formats, while maintaining ease of implementation through a unified processing architecture that adapts to different data characteristics.
Data Source
AI summary
Techniques are disclosed relating to text sanitization. Given textual data, a computer system identifies tokens predicted to constitute sensitive information. Multi-field data structures (e.g., triplets) are generated for the identified tokens that include questions, answers, and corresponding context. These data structures are supplied to a pre-trained multiple-choice question (MCQ) reading comprehension model. The model outputs, for each data structure, a probability that the question and answer for a given data structure, provided the context, is accurate. A post-processing module can then rank probabilities in this set of probabilities and select the multi-field data structure with the highest probability (in some cases, a programmable threshold must also be met). The selected multi-field data structure is then used to select category information to be used in sanitizing the textual data. In this manner, a piece of sensitive data may be replaced by a label that helps retain interpretability of the sanitized text.