Training Broker for Privacy-Preserving Distributed ML Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Customers are hesitant to provide private or sensitive information for machine learning model training due to privacy concerns, which hinders the creation of accurate models and complicates data sharing among distributed customers.
Innovation Solution
A computer-implemented method using a training broker to obtain anonymized data from multiple sources, processed through a data science schema that filters out sensitive information, allowing for machine learning model training without exposing private data, thereby maintaining customer privacy and enabling distributed model training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If production data is used to train machine learning models, then model accuracy and performance are improved, but customer privacy and data security are compromised
Solution Approach 1:
The patent introduces a training broker as an intermediary between customers and the machine learning vendor. The broker obtains anonymized data from customers, trains models using this data, and provides results back to customers without the vendor ever accessing the raw production data. This mediator resolves the contradiction by enabling model training with real data while preventing direct vendor access to sensitive information.
Solution Approach 2:
The patent extracts only the necessary information for model training from the production data through anonymization processes. The training broker requests specific anonymized data elements from customers' production systems, separating the useful training information from the sensitive personal identifiers. This extraction allows model training to proceed while removing privacy risks.
2Productivity
If customers provide their data to the vendor for model training, then model performance is improved, but data privacy and information security are compromised
Solution Approach 1:
The training broker serves as a trusted intermediary that customers can share data with without risking their privacy. The broker handles data transfer and model training operations, allowing customers to participate in collaborative model training while maintaining control over their data privacy. This resolves the contradiction by enabling efficient training through data sharing while preventing information loss to unauthorized parties.
3Measurement precision
If real production data is used instead of manufactured test data, then model accuracy is improved, but data privacy concerns arise
Solution Approach 1:
The patent extracts and removes sensitive information elements from production data through anonymization techniques. The training broker requests only anonymized versions of the data that retain the necessary patterns and relationships for accurate model training while stripping out personally identifiable information. This extraction approach enables use of real production data without the complexity of implementing comprehensive privacy protection mechanisms at the customer side.
Data Source
AI summary
Computer-implemented systems and computer-implemented methods include the following. A request to train a machine-learning (ML) model is received at a training broker. Anonymized data for training the model is obtained by the training broker from each individual data source of a plurality of data sources. The anonymized data is accessed through a data science schema being provided by anonymization of sensitive information of production data from each individual data source. Access to the anonymized data is provided to a data vendor for training the ML model using the anonymized data.