Multi-party model training via aggregation system
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Organizations face challenges in collaborating on machine learning model training due to concerns over data privacy, proprietary information, and regulatory compliance, as existing solutions require data sharing which is often not feasible or accessible.
Innovation Solution
A system and method for multi-party model training that allows data owners to collaborate on generating a model without sharing their actual data, using an aggregation system to manage model versions and access controls, ensuring that only approved updates are incorporated into the aggregated model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If data is shared between organizations for collaborative model training, then model accuracy and collaboration effectiveness are improved, but data privacy and security are compromised
Solution Approach 1:
The patent introduces a trusted third-party intermediary system that coordinates multi-party model training without requiring direct data sharing between organizations. This intermediary manages the training process, aggregates model updates, and ensures that raw data remains isolated within each organization's secure environment, thus achieving collaboration while protecting data privacy.
Solution Approach 2:
The patent segments the model training process into separate components that can be executed independently by different organizations. Each organization trains local model instances on their own data, and only model parameters or gradients are shared, not the raw data itself. This segmentation allows collaborative training while maintaining data isolation.
2Measurement precision
If data is shared between organizations for collaborative model training, then model accuracy is improved, but compliance with regulatory standards deteriorates
Solution Approach 1:
The intermediary system implements regulatory compliance mechanisms while facilitating model training. It ensures that data sharing policies, access controls, and audit trails are maintained according to regulatory requirements, allowing organizations to collaborate on model training without violating data protection regulations.
3Object-affected harmful factors
If data is not shared between organizations, then data privacy and security are maintained, but collaboration effectiveness deteriorates
Solution Approach 1:
The patent uses model copying where each organization receives a copy of the base model or model checkpoints, trains locally on their own data, and contributes updates back to the central system. This copying approach enables collaboration without requiring organizations to share their actual data, maintaining privacy while achieving collaborative benefits.
4Object-affected harmful factors
If access controls are implemented to protect data privacy, then data security is improved, but system complexity increases
Solution Approach 1:
The intermediary system centralizes access control management, simplifying the overall system architecture. Instead of each organization implementing complex access controls independently, the intermediary manages authentication, authorization, and data flow control, reducing system complexity while maintaining strong security.
Data Source
AI summary
Methods and systems for collaboration between two or more parties to generate a model based on data from the parties without data to the another party. A system includes an aggregation system having computer storage devices configured to store a model and a plurality versions of an aggregated model, and instructions, and one or more processors configured to execute the plurality of computer readable instructions to, iteratively, receive a first updated model from a first system controlled by a first party and a second updated model from a second system controlled by a second party, determine changes from the first updated model and the second updated model to include in an aggregated model, communicate a version of the aggregated model to the first system and a version of the aggregated model to the second system, and store a final aggregated model.