Privacy-Enhanced Prompt Processing for Large Language Models

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Large generative AI models (LXMs) often inadvertently disclose privacy information when responding to user prompts, particularly in non-private locations where the responses can be viewed by others.

Innovation Solution

The system processes user prompts to recognize potential privacy information and generates privacy-enhanced prompts to avoid disclosure. This involves selecting a locally-hosted LXM, generating contextualized prompts, and reviewing responses to ensure privacy is maintained.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If LXMs are used to provide responses to user prompts, then the functionality and understanding capabilities are improved, but privacy information may be inadvertently disclosed

Engineering Contradiction:

ImprovefunctionalityVSAvoidprivacy disclosure

Core Design Contradiction:

Adaptability or versatilityVSObject-affected harmful factors

Solution Approach 1:

A trained model is introduced as an intermediary between the user prompt and the LXM. This intermediary processes the user prompt to recognize privacy information and generates a privacy-enhanced prompt that is then submitted to the LXM, preventing direct exposure of private information while maintaining the LXM's functionality

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The system performs preliminary processing of the user prompt before it reaches the LXM. The trained model analyzes the prompt in advance to identify privacy-sensitive content and transforms it into a privacy-enhanced version, preventing privacy disclosure before it can occur

Inventive Principle:

Principle #10Preliminary action

2Object-affected harmful factors

If privacy-enhanced prompts are generated and processed, then privacy protection is improved, but the complexity of the system increases

Engineering Contradiction:

Improveprivacy protectionVSAvoidsystem complexity

Core Design Contradiction:

Object-affected harmful factorsVSDevice complexity

Solution Approach 1:

The trained model serves multiple functions: it processes user prompts to recognize privacy information, determines whether privacy enhancement is needed, generates privacy-enhanced prompts, and can also select appropriate LXM models. This multi-functionality reduces the need for separate dedicated components for each task

Inventive Principle:

Principle #6Universality (Multi-functionality)

3Object-affected harmful factors

If locally-hosted LXMs are selected for privacy-sensitive prompts, then privacy protection is improved, but the processing speed and response time may decrease

Engineering Contradiction:

Improveprivacy protectionVSAvoidresponse time

Core Design Contradiction:

Object-affected harmful factorsVSSpeed

Solution Approach 1:

The system dynamically selects between cloud-based and locally-hosted LXM models based on the privacy sensitivity of the prompt. For non-sensitive prompts, faster cloud-based models are used; for sensitive prompts, locally-hosted models are selected to ensure privacy protection, optimizing the balance between speed and privacy

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS20250131121A1Privacy Enhanced Language Model Prompt

Publication Date: 2025.04.24 QUALCOMM INC

AI summary

Various embodiments include systems and methods for generating a prompt for a large generative AI model (LXM). A computing device may be configured to receive a user prompt, process the received user prompt to recognize whether the prompt includes privacy information or will cause an LXM to provide a response that will reveal privacy information; and use the LXM to provide a response to the user prompt in a manner that will avoid disclosure of privacy information.