LLM Input Redaction With Semantic Remapping for Confidential Text
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for interacting with large language models fail to prevent confidential data leakage while maintaining the quality of generated text, posing risks to individuals and organizations, especially in regulated sectors.
Innovation Solution
A data processing method and system that converts confidential data into de-identified common words using a semantic mapping model, processes the text locally, and then unmaps the results to maintain natural language structure and confidentiality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If public key encryption technique is used to encrypt information, then eavesdropping attacks during transmission are prevented, but confidential data can still be provided to the service server after decryption
Solution Approach 1:
The patent extracts and removes confidential data from the text input before sending it to the large language model. By identifying and extracting sensitive information (such as personal identifiers, financial data, etc.) and replacing it with placeholders or generic terms, the system prevents confidential data from reaching the external service server, thereby solving the contradiction between transmission security and data protection.
Solution Approach 2:
The patent introduces an intermediary processing layer between the user and the large language model service. This intermediary system performs data redaction, masking, and anonymization before the text is sent to the external model, acting as a protective barrier that prevents direct exposure of confidential data while maintaining the functional integrity of the interaction.
2Reliability
If ordinary string replacement method is applied to prevent data leakage, then confidential information is replaced with alias name strings, but the quality of text context processing and generated results is significantly affected
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different text elements. Instead of uniform replacement, the system identifies specific confidential data patterns (names, dates, numbers, etc.) and applies targeted redaction only to those elements while preserving the natural flow and quality of the remaining text. This localized approach maintains text context processing quality while preventing data leakage.
Solution Approach 2:
The patent changes the parameters of text processing by using context-aware replacement strategies. Rather than simple string substitution, the system analyzes the semantic context and adjusts the replacement approach accordingly, using placeholders, generic terms, or synthesized alternatives that maintain the textual structure and meaning. This parameter adjustment ensures high-quality text processing while achieving effective data protection.
Data Source
AI summary
A data processing method for interacting with large language models, comprising obtaining an original text, analyzing the original text and determining confidential data items in the original text, generating a semantic model representing the relations across the items, converting the confidential data items in the original text into a common word to generate a secure input text, inputting the secure input text to a large language model for processing, receiving a reply text generated by the large language model based on the secure input text, and converting the reply text into a target reply text corresponding to the original text according to the common word and the semantic model.