Encrypted Predictive Model Parameters for Secure Private Data Access

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing data encryption and access authorization schemes are inadequate for protecting predictive models trained on private user data, as they fail to provide transparent access to non-private model components while safeguarding private information.

Innovation Solution

A predictive application that securely stores and executes computer-implemented predictive models by encrypting learned parameters and using a content key, while keeping execution code unencrypted, ensuring only authorized processes can access the encrypted parameters.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional data encryption schemes are applied to protect predictive models trained on private data, then security of private data is improved, but transparent access to non-private model components is lost

Engineering Contradiction:

Improvesecurity of private dataVSAvoidtransparent access to non-private model components

Core Design Contradiction:

ReliabilityVSEase of operation

Solution Approach 1:

The predictive model is segmented into two distinct parts: encrypted learned parameters (stored in encrypted model file) and unencrypted execution code (stored in unencrypted model file). This segmentation allows the execution code to remain accessible for development and maintenance while the learned parameters containing private data information are protected through encryption.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Different security levels are applied to different parts of the model: the execution code is kept unencrypted with high accessibility for developers, while only the learned parameters are encrypted with high security protection. This local quality approach ensures that transparency is maintained where needed while protection is applied where necessary.

Inventive Principle:

Principle #3Local quality

2Reliability

If learned parameters are encrypted to protect private data, then unauthorized access to private information is prevented, but model execution and development are restricted

Engineering Contradiction:

Improveprotection of private informationVSAvoidmodel execution and development flexibility

Core Design Contradiction:

ReliabilityVSAdaptability or versatility

Solution Approach 1:

The model is divided into encrypted learned parameters and unencrypted execution code, allowing the system to maintain both security and flexibility. The unencrypted execution code enables model developers to execute, test, and maintain the model without decryption, while the encrypted parameters protect private data.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system introduces an intermediary mechanism where the unencrypted execution code acts as a mediator that can process inputs and generate outputs without directly accessing the encrypted learned parameters in plaintext, thus maintaining both security and operational flexibility.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Ease of manufacture

If all model components are kept unencrypted for developer access, then ease of development is improved, but private user data becomes vulnerable to unauthorized access

Engineering Contradiction:

Improveease of model developmentVSAvoidvulnerability of private user data

Core Design Contradiction:

Ease of manufactureVSObject-affected harmful factors

Solution Approach 1:

The model components are segmented into learned parameters (containing private data) and execution code (needed for development). By encrypting only the learned parameters while keeping execution code unencrypted, the system enables continued ease of development without exposing private user data to unauthorized access.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Encryption is applied locally only to the learned parameters that contain private data information, while the execution code remains unencrypted for developer accessibility. This selective application of encryption maintains development ease while protecting against harmful factors.

Inventive Principle:

Principle #3Local quality

Data Source

PatentUS11514366B2System for content encryption for predictive models trained on private data

Publication Date: 2022.11.29 MICROSOFT TECHNOLOGY LICENSING LLC

AI summary

A computing system receives decrypts learned parameters of a computer-implemented predictive model stored in a computer-readable, non-volatile data store upon instantiation of a model process in volatile memory of the computing system. The predictive model is parameterized by learned parameters that are based upon private data of a user, wherein the learned parameters of the predictive model have been encrypted subsequent to training of the predictive model based upon the private data of the user. The computing system receives a request for a prediction associated with the user. Upon authorizing the request, the computing identifies the predictive model. The computing system, by way of the model process, generates the prediction based upon the request and the predictive model. The computing system outputs the prediction to a computing device associated with the user.