Split Generative Model Inference for Privacy-Safe Cloud Content Generation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The transmission of sensitive information in content generation instructions to cloud-based generative models poses a risk of data leakage, compromising user privacy.

Innovation Solution

Deploying a portion of the generative model on the user's device to process content generation instructions, generating an intermediate result that is difficult to deduce, and sending this result to the cloud for final content generation, thereby preventing direct exposure of sensitive information.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If the content generation instruction is sent to the cloud-based generative model, then the model can generate target content, but the sensitive information of the user may be leaked

Engineering Contradiction:
Improvecontent generation capabilityVSAvoidinformation security risk
Core Design Contradiction:
ProductivityVSObject-affected harmful factors

Solution Approach 1:

The generative model is divided into two parts: the first module (interpreter) is deployed on the user end to process the content generation instruction and output an intermediate result, while the second module (generator) is deployed on the cloud to generate the final target content based on the intermediate result. This segmentation ensures that sensitive information in the original instruction does not directly reach the cloud, thus preventing information leakage while maintaining content generation capability.

Inventive Principle:
Principle #1Segmentation

2Object-affected harmful factors

If the whole model is deployed on the user end, then information security is improved, but the computing power requirement increases

Engineering Contradiction:
Improveinformation securityVSAvoidcomputing power requirement
Core Design Contradiction:
Object-affected harmful factorsVSPower

Solution Approach 1:

The model is segmented such that the computationally intensive first module runs locally on the user end device, while the second module runs on the cloud. This distribution optimizes the balance between local processing requirements and cloud resources, reducing the burden on user end computing power while still maintaining information security through local processing of sensitive data.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Instead of requiring the user end to run the entire model, only the first module (interpreter) is executed locally to generate an intermediate result. The second module (generator) is executed on the cloud. This partial action approach satisfies the information security requirement without imposing excessive computing power demands on the user end device.

Inventive Principle:
Principle #16Partial or excessive action

3Object-affected harmful factors

If the first module processes the content generation instruction locally, then sensitive information is protected, but the device complexity increases

Engineering Contradiction:
Improvesensitive information protectionVSAvoidsystem deployment complexity
Core Design Contradiction:
Object-affected harmful factorsVSDevice complexity

Solution Approach 1:

The system is segmented into two deployable modules with clear functional boundaries. The first module (interpreter) handles local processing of content generation instructions, while the second module (generator) handles cloud-based content generation. This segmentation, while introducing some complexity, provides a structured and manageable approach to protecting sensitive information through localized processing.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4715650A1Method and system for generating content by using model, and cluster
Publication Date: 2026.03.25 HUAWEI CLOUD COMPUTING TECHNOLOGIES CO LTD
  • EP4715650A1 patent drawingFigure 1
  • EP4715650A1 patent drawingFigure 2
  • EP4715650A1 patent drawingFigure 3A~3B

AI summary

A method and system for generating content by using a model, and a cluster are provided and may be applied to the field of cloud computing technologies. The method includes: A client receives a content generation instruction sent by a user, where the content generation instruction instructs a model to generate target content based on an intention of the user. The client inputs the content generation instruction into a first module, to enable the first module to output an intermediate result of generating the target content by the model. A serving end receives the intermediate result sent by the client, and inputs the intermediate result into a second module, to enable the second module to output the target content. The serving end sends the target content to the client. The method can prevent a cloud from obtaining sensitive information of the user in the content generation instruction, and protecting information security of the user.