LLM Data Anonymization via Intermediary Obfuscation Pipeline
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems fail to ensure data privacy for users interacting with large language models (LLMs), leading to GDPR concerns and user distrust due to the potential misuse of sensitive data during training and utilization.
Innovation Solution
Implement a system for LLM-agnostic data anonymization that includes authentication, interception, parsing, detection, obfuscation, and deobfuscation engines to protect sensitive information, ensuring secure interactions with LLMs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If users send sensitive data to LLMs for processing, then LLM analytics and training benefits are improved, but data privacy and security deteriorate
Solution Approach 1:
The patent introduces an intermediary anonymization system that sits between the user's sensitive data and the LLM service. This intermediary layer detects, obfuscates, and redacts confidential information before data reaches the LLM, while still allowing the LLM to process the anonymized data for analytics and training purposes. The system acts as a mediator that enables data utilization while protecting privacy.
Solution Approach 2:
The patent applies preliminary action by performing data anonymization before the data is sent to the LLM service. The system detects confidential information, generates obfuscated versions, and redacts the original data in advance. This preliminary processing ensures that sensitive data is protected before it enters the LLM pipeline, preventing potential privacy breaches during training and analytics.
2Reliability
If data anonymization is implemented before LLM processing, then data privacy is improved, but system complexity increases
Solution Approach 1:
The patent segments the anonymization system into distinct functional modules: detection engine, obfuscation engine, and redaction engine. Each module performs a specific function in the anonymization pipeline. The detection engine identifies confidential information, the obfuscation engine generates protected versions, and the redaction engine replaces original data with obfuscated versions. This segmentation makes the complex anonymization process more manageable and maintainable.
Solution Approach 2:
The patent creates a universal anonymization system that can handle multiple types of confidential information and integrates with various LLM services. The system uses machine learning classifiers that can be trained on different data types and domains, making it adaptable to various privacy protection scenarios. This multi-functional approach reduces the need for separate specialized systems for different anonymization tasks.
3Measurement precision
If machine learning classifiers are trained on annotated datasets, then detection accuracy is improved, but data processing time increases
Solution Approach 1:
The patent applies partial action by training machine learning classifiers on annotated datasets that represent only the most critical and common types of confidential information. Rather than training on all possible data types, the system focuses on high-impact information categories, achieving sufficient detection accuracy for the most important privacy protection tasks while reducing training time and computational resources required.
Data Source
AI summary
Systems and methods for large language model (LLM)-agnostic data anonymization. Data anonymization includes data obfuscation (and data deobfuscation) to protect confidential information a user is going to send to an LLM service or application programming interface (API). Encryption can be used for data obfuscation and particularly, for securing data from unauthorized access. Likewise, decryption can be used for data de-obfuscation.