Sensitive Data Obfuscation Before Generative AI Ingestion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large Language Models (LLMs) retain sensitive user data during interactions, posing privacy and security risks due to their inability to differentiate and protect sensitive information within user inputs.
Innovation Solution
A system is developed to identify sensitive data within user requests, translate and compare fields across languages to detect untranslated data, and obfuscate or delete sensitive information before transmitting it to AI engines, with options for user interaction in the obfuscation process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If LLMs process user inputs to generate tailored responses, then the interaction quality and personalization are improved, but sensitive user data is retained and exposed to privacy and security risks
Solution Approach 1:
The system segments user input into sensitive and non-sensitive components by translating the request to multiple languages and comparing fields. Fields that remain untranslated are identified as sensitive data and separated from the rest of the request for obfuscation, allowing the non-sensitive portions to be processed normally for maintaining interaction quality.
Solution Approach 2:
The patent introduces an intermediary processing layer between user input and the LLM that acts as a mediator. This layer translates requests to multiple languages, identifies sensitive fields through comparison, obfuscates sensitive data, and then transmits the processed request to the AI engine, thereby protecting privacy while preserving interaction functionality.
2Measurement precision
If sensitive data is transmitted to AI engines for processing, then the AI can provide accurate and contextually relevant responses, but security and compliance requirements are violated
Solution Approach 1:
The system performs preliminary action by obfuscating sensitive data in user requests before they are transmitted to the AI engine. The translation and comparison process identifies sensitive fields in advance, allowing obfuscation to occur prior to AI processing, thus ensuring compliance requirements are met while the AI still receives sufficient context for accurate responses.
3Object-affected harmful factors
If all user data is obfuscated before transmission, then privacy protection is improved, but the AI engine loses contextual understanding and response quality deteriorates
Solution Approach 1:
The system applies local quality by differentiating between sensitive and non-sensitive portions of user input. Only the identified sensitive fields are obfuscated, while non-sensitive portions retain their original form and contextual meaning. This selective approach protects privacy where needed while preserving contextual information for the AI engine to generate accurate responses.
Data Source
AI summary
An approach is provided that identifies a sensitive data in a request to an artificial intelligence (AI) engine. In one embodiment, the identifying further includes: translating the request from one natural language to other natural languages, thus creating translated requests, comparing fields in the first natural language to fields in the other natural languages with the comparing resulting in some untranslated fields that are treated as the sensitive data. The approach further includes creating an obfuscated request by obfuscating the sensitive data identified in the request before the obfuscated request is transmitted to the AI engine.