Cross-Domain Machine Learning With MPC for Privacy-Preserving Personalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models trained across multiple sources risk leaking private or sensitive data, and the use of third-party cookies compromises user privacy and security, especially as they are deprecated.
Innovation Solution
Implement secure multi-party computation (MPC) techniques to train machine learning models using encrypted user profiles and contextual signals, ensuring that no single party can access complete user data in cleartext, enabling the selection of digital components based on predicted performance measures without revealing individual user identities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are trained using data from multiple sources, then model accuracy and performance are improved, but user privacy and data security are compromised due to potential data leakage
Solution Approach 1:
The patent segments user data into multiple encrypted shares distributed across different computing systems. Each system holds only a portion of the encrypted data, making it impossible for any single system to access complete user profiles. This segmentation enables collaborative model training while preventing data concentration and potential leakage.
Solution Approach 2:
The patent introduces an encrypted secret share mechanism as an intermediary between data sources and the machine learning model. This intermediary layer allows multiple parties to contribute to model training without directly sharing sensitive user data, thus preventing data leakage while maintaining model accuracy.
2Adaptability or versatility
If third-party cookies are used for tracking user behavior, then targeted advertising and user experience personalization are improved, but user privacy and security are compromised
Solution Approach 1:
The patent extracts the essential functionality of third-party cookies (user behavior tracking for personalization) while removing the privacy-invasive aspect. By using encrypted secret shares and multi-party computation, the system achieves personalization capabilities without requiring traditional cookies that compromise user privacy.
Solution Approach 2:
The patent changes the fundamental parameters of how user data is stored and accessed. Instead of using clear-text cookies that can be easily accessed and tracked, the system uses encrypted secret shares with cryptographic protection, fundamentally altering the security and privacy parameters while maintaining personalization functionality.
3Measurement precision
If complete user profiles are transmitted to content platforms, then digital component selection accuracy is improved, but bandwidth consumption and data transmission time increase
Solution Approach 1:
The patent applies partial action by transmitting only encrypted secret shares of user profiles rather than complete unencrypted profiles. This partial transmission approach maintains the necessary information for accurate component selection while significantly reducing the amount of data transmitted and bandwidth consumed.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
This document describes a secure machine learning platform. In some aspects, a method includes transmitting by the application to the machine learning platform, a set of data including a user profile, one or more characteristics of a digital component, contextual signals, model identifier, and data indicating a type of event. The application receives a request generated based on the computer-readable instructions to upload a user profile of a user of the client device to a machine learning platform. The computer-readable instructions initiate the request in response to detecting an occurrence of the event with the digital component. In response to the request, the application can obtain the user profile request data element that includes a model identifier for a machine learning model and one or more characteristics of at least one of the digital component or the first content page.