Dynamic Network Analysis for LLMs with Format-Preserving Encryption
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large and complex digital networks face challenges in information exchange and resource sharing due to interconnectedness, and existing encryption methods hinder the use of large language models by altering information formats, making it difficult for them to draw accurate conclusions.
Innovation Solution
A method and system for encrypting information using word-key pairs that maintain referential integrity and format preservation, allowing a large language model to understand and respond consistently while protecting sensitive data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If standard encryption is used to protect sensitive information, then security is improved, but the ability of large language models to understand and process the information deteriorates
Solution Approach 1:
The patent introduces format-preserving encryption as an intermediary mechanism that encrypts sensitive information while maintaining the original format and structure. This allows the encryption to serve as a mediator between security requirements and model processing capabilities, enabling the LLM to process encrypted data that retains recognizable patterns and relationships.
Solution Approach 2:
The patent changes the encryption approach from standard encryption to format-preserving encryption, which modifies the encryption parameters to maintain the original data format. This parameter change allows the encrypted data to retain structural characteristics that LLMs can process, such as grammatical patterns and semantic relationships, while still providing security.
2Object-affected harmful factors
If encryption replaces words and phrases to protect information, then security is improved, but the consistency and accuracy of large language model inference deteriorates
Solution Approach 1:
The patent applies format-preserving encryption that maintains the original format of words and phrases while encrypting them. This parameter change ensures that encrypted text retains the structural and contextual information needed for consistent LLM inference, allowing the model to maintain reliable processing despite the encryption.
Solution Approach 2:
Instead of encrypting data and then attempting to process it (which causes inference accuracy to deteriorate), the patent inverts the approach by using format-preserving encryption that inherently maintains processability. The encryption is designed from the outset to preserve the formats and structures that LLMs rely on for accurate inference.
3Object-affected harmful factors
If complex encryption methods are used to secure network information, then security is improved, but the complexity of information processing and analysis increases
Solution Approach 1:
The patent changes the encryption parameters to use format-preserving encryption, which simplifies the processing complexity by maintaining the original data formats. This allows standard LLM processing techniques to work directly with encrypted data without requiring complex decryption or specialized processing, thus reducing overall system complexity while maintaining security.
Data Source
AI summary
A method of determining a natural language output regarding a digital network using a large language model (“LLM”) can include formulating a desired output dependent upon information associated with the digital network; providing, to the LLM, the information associated with the digital network and a first prompt requesting the LLM to generate a query dependent upon the information and the desired output; receiving, from the LLM, the query dependent upon the information and the desired output; determining, dependent upon a graph database, a response to the query with the graph database being representative of at least a portion of the digital network; providing, to the LLM, the response and a second prompt requesting the LLM to generate the natural language output dependent upon the response; and receiving, from the LLM, the natural language output dependent upon the response and associated with the digital network.