Confidential Computing for Large Model Data Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for using large models pose privacy concerns due to data leakage risks, especially with centralized desensitization services, which can access sensitive user information and authorization details, and require significant hardware deployment or local services.
Innovation Solution
A secure and trusted method using confidential computing, where a desensitization service is deployed in a confidential computing environment, establishing secure connections with user terminals and large models, desensitizing data within the environment, and using trusted sealing technology to protect user data and authorization information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If centralized desensitization services are used to protect data privacy, then data desensitization is achieved, but the service can still access sensitive user information and authorization details
Solution Approach 1:
The patent introduces a confidential computing environment as an intermediary layer between the user terminal and the centralized desensitization service. This TEE acts as a trusted mediator that performs desensitization operations while keeping all data, including sensitive user information and authorization details, isolated from the service's main processing memory. The intermediary hardware enclave ensures that even though the service can access data for processing, it cannot leak sensitive information.
Solution Approach 2:
The patent segments the computing environment into separate memory spaces: a trusted execution environment (TEE) for desensitization operations and a non-trusted service environment for business logic. This segmentation physically separates sensitive data handling from the service's accessible memory, preventing the service from accessing or leaking user authentication information while maintaining necessary desensitization functionality.
2Reliability
If local deployment of de-anonymization services is implemented, then data privacy is protected, but implementation cost and hardware requirements increase
Solution Approach 1:
The patent uses a cloud-based confidential computing environment as an intermediary service that users can access without local deployment. Instead of requiring users to install and maintain complex de-anonymization services locally, the system provides this functionality as a remote service through the TEE, significantly reducing local hardware complexity while maintaining privacy protection.
Solution Approach 2:
The patent replicates the desensitization functionality in a remote confidential computing environment rather than requiring duplication on every user terminal. This allows multiple users to access the same trusted desensitization service without each user needing to deploy the full service stack locally, reducing overall system complexity.
3Reliability
If local deployment of large models is implemented, then data privacy is protected, but hardware resources are required that ordinary users lack
Solution Approach 1:
The patent introduces a cloud-based confidential computing environment as an intermediary that hosts the large model and performs desensitization operations remotely. This eliminates the need for users to possess high-performance hardware for local model deployment while maintaining privacy protection through the TEE's secure processing of data before it reaches the model.
Solution Approach 2:
The patent moves the computationally intensive operations (large model inference and desensitization) from the user terminal to a remote confidential computing environment. This dimensional shift from local to remote execution allows users to access powerful computing resources without requiring them to own or deploy such hardware locally.
Data Source
AI summary
Provided is a secure and trusted use method and system for a large model based on confidential computing. In the method, a desensitization service party starts a service in a confidential computing environment; the confidential computing environment establishes a secure connection with a user terminal and the large model separately; the service receives, in the confidential computing environment, original data uploaded by the user terminal, and desensitizes the original data to obtain desensitized data; and the service sends the desensitized data to the large model based on the secure connection between the confidential computing environment and the large model. In this scheme, a third party is used to desensitize the data before utilizing the large model, and a targeted solution is provided to address the issue of third-party data leakage. Confidential computing is applied to the usage scenario of the large model, with the third-party service performing desensitization.