Language Model Key-Value Cache Sharing Without Privacy Exposure
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The secure sharing of key-value caches in language models poses a security risk due to the potential exposure of protected information, compromising user privacy, while disabling reuse of key-values entirely reduces interaction efficiency.
Innovation Solution
A method and apparatus that recognize protected and non-protected information in request content, adding key-values corresponding to non-protected information to a shared cache while excluding protected information, ensuring secure and efficient reuse.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If key-value cache sharing is enabled to improve interaction efficiency, then computing time and resources are saved, but protected information may be exposed compromising user privacy
Solution Approach 1:
The patent segments the key-value cache into protected and non-protected portions by identifying and masking sensitive information in request content before caching. This allows partial reuse of cache entries while preventing exposure of private data, thus resolving the contradiction between efficiency improvement and privacy protection.
Solution Approach 2:
The patent applies different quality treatments to different parts of the request content: non-protected information is cached for reuse to improve efficiency, while protected information is masked or excluded from the cache to prevent privacy exposure. This local differentiation resolves the contradiction by optimizing each region according to its sensitivity level.
2Object-affected harmful factors
If key-value cache sharing is disabled to protect user privacy, then protected information security is improved, but interaction efficiency deteriorates due to loss of reuse capability
Solution Approach 1:
The patent implements partial caching by storing only the non-protected portions of request content in the key-value cache while excluding or masking protected information. This partial action maintains security for sensitive data while still enabling efficiency improvements through reuse of non-sensitive cached entries.
Solution Approach 2:
The patent introduces an information recognition and masking mechanism as an intermediary between the request content and the key-value cache. This intermediary identifies protected information, applies appropriate masking or exclusion, and allows safe caching of remaining non-protected content, thus enabling both security and efficiency.
3Productivity
If all key-values are added to shared cache to maximize reuse, then interaction efficiency is improved, but security risk increases due to potential exposure of protected information
Solution Approach 1:
The patent performs preliminary identification and masking of protected information in request content before the caching operation. This preliminary action ensures that only safe, non-protected information is added to the shared cache, maximizing reuse efficiency while preventing security risks from the outset.
Solution Approach 2:
The patent extracts and removes protected information from the request content before adding the remaining non-protected portions to the shared key-value cache. This extraction ensures that only safe data is cached for reuse, maintaining both high efficiency and security reliability.
Data Source
AI summary
According to embodiments of the disclosure, a method, an apparatus, a device and a medium for securely sharing key-value cache of a language model are provided. The method includes: obtaining request content input to a language model; recognizing protected information and non-protected information in the request content; determining, for a first part in the non-protected information whose matched key-value is absent from a shared key-value cache of the language model, a key-value corresponding to the first part and adding the key-value to the shared key-value cache, where a key-value of the protected information is not added to the shared key-value cache.