Secure Container Prediction Model for Privacy-Preserving User Request Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in securely processing user requests across multiple parties while maintaining data privacy, as sensitive data needs to be transmitted and trained models are often created on a feature data provider, leading to risks and performance overheads due to long-distance data transmission.
Innovation Solution
The method involves creating a secure container using Software Guard Extensions technology on the data user's local storage space, where an initial prediction model is trained with encrypted feature and labeled result samples, and the encrypted feature sample is transmitted through a ciphertext transmission path, allowing the target prediction model to process user requests without exposing sensitive data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the initial prediction model is trained on the feature data provider's system, then the model can be created using the provider's computational resources, but sensitive data must be transmitted long-distance which increases data exposure risk and transmission overhead
Solution Approach 1:
Instead of transmitting data to the model training location (conventional approach), the patent inverts the process by bringing the model training capability to the data user's location through encrypted model parameters. The initial prediction model is trained on the feature data provider's system, but only encrypted model parameters are transmitted to the data user, who then decrypts and uses them locally, eliminating the need for long-distance sensitive data transmission.
Solution Approach 2:
The patent creates a copy of the model functionality through encrypted model parameters that can be deployed locally at the data user's system. Rather than moving the actual sensitive data or the full model execution environment, a cryptographic copy (encrypted parameters) is transmitted that reproduces the model's predictive capability locally, reducing transmission distance and exposure risk.
2Adaptability or versatility
If sensitive data is transmitted through the network for model training, then the model can be trained with comprehensive data from multiple parties, but the data exposure risk increases during transmission
Solution Approach 1:
The patent changes the state of the model parameters from plaintext to encrypted form during transmission and storage. The initial prediction model is trained with comprehensive data from multiple parties, but the model parameters are encrypted before being transmitted to the data user. This parameter transformation maintains the comprehensiveness of model training while eliminating data exposure risk during network transmission.
Solution Approach 2:
Encryption serves as an intermediary mechanism between the model training process and the data user's system. The encrypted model parameters act as a mediator that carries the trained model's knowledge without exposing the underlying sensitive data. This intermediary layer allows comprehensive model training while protecting sensitive information during transmission and deployment.
3Ease of manufacture
If the prediction model is deployed remotely on the feature data provider's system, then centralized model management is achieved, but performance overhead increases due to long-distance data transmission between user and model
Solution Approach 1:
The patent extracts the essential predictive capability from the centralized model and places it locally at the data user's system in the form of encrypted model parameters. While centralized model management is maintained for model creation and updates, the actual inference process is extracted and executed locally, eliminating long-distance data transmission during prediction and reducing time loss.
Solution Approach 2:
The model training and parameter encryption are performed in advance (preliminarily) on the feature data provider's system. The encrypted model parameters are prepared beforehand and deployed to the data user's local system. This preliminary action allows centralized model management while enabling fast local inference without real-time transmission delays.
4Reliability
If encrypted model parameters are transmitted to the data user's system, then data security is improved by keeping sensitive data local, but the device complexity increases due to encryption and decryption operations
Solution Approach 1:
The patent replaces complex security mechanisms (such as secure hardware enclaves, trusted execution environments, or distributed secure computing infrastructure) with cryptographic substitution. Instead of relying on complex mechanical or architectural security systems, the solution uses encryption and decryption operations to achieve the same security goals with simpler implementation. The encrypted model parameters can be securely transmitted and stored using standard cryptographic protocols, reducing device complexity compared to alternative secure computation approaches.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
Embodiments of the present disclosure provide a method, apparatus and electronic device for processing a user request, and a computer readable storage medium. A particular implementation of the method includes: receiving a user request; sending the user request to a target prediction model stored in a secure container, where the secure container is created in a local storage space by using Software Guard Extensions technology, and the target prediction model is obtained by training an initial prediction model with an encrypted feature sample and a labeled result sample corresponding to the encrypted feature sample, and the encrypted feature sample is transmitted by a feature data provider through a ciphertext transmission path established between the feature data provider and the secure container; and receiving a prediction result output by the target prediction model.