Distributed ML Proxy Modeling for Private Data Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Researchers face challenges in accessing and combining distributed, private data sets for machine learning due to privacy restrictions, leading to incomplete datasets and loss of valuable information during de-identification, which existing technologies fail to address effectively.
Innovation Solution
A distributed machine learning system that generates trained proxy models and shares learned information without exposing private data, using private data servers to create trained models and distribute proxy data representative of the original data's characteristics, allowing integration into an aggregated model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If de-identification is performed to remove references to specific patients before providing data to researchers, then privacy protection is improved, but information completeness deteriorates
Solution Approach 1:
The patent extracts only the necessary statistical characteristics and patterns from private data sets, separating the valuable information (statistical properties) from the sensitive content (identifiable information). This allows researchers to access statistical insights without exposing private individual records, resolving the contradiction between privacy protection and information completeness.
Solution Approach 2:
The system introduces an intermediary processing layer that transforms private data into statistical summaries and models. This intermediary layer acts as a buffer between the private data sources and researchers, enabling information exchange while maintaining privacy boundaries. The intermediary extracts and transmits only statistical characteristics, not raw private data.
2Measurement precision
If distributed private data sets are combined to build comprehensive machine learning models, then model accuracy is improved, but data access difficulty worsens
Solution Approach 1:
The patent extracts statistical characteristics from distributed private data sets without requiring direct access to the underlying raw data. By extracting and transmitting only the necessary statistical properties (means, variances, correlations, etc.), the system enables model building across distributed data sources while avoiding the access difficulties associated with centralized data collection.
Solution Approach 2:
The system creates a universal interface for accessing distributed data through statistical summaries. Instead of requiring different access methods for different data sources, the patent establishes a unified approach where all private data sets are accessed through their statistical characteristics, enabling comprehensive model building across diverse distributed sources.
3Adaptability or versatility
If statistical models are merged to create final multi-dimensional classification spaces, then predictive capability is improved, but assumption of data access authorization worsens the privacy problem
Solution Approach 1:
The patent inverts the traditional approach by not merging raw data sets to create classification spaces, but rather merging statistical models that have already been trained on distributed private data. Instead of centralizing data and then modeling, the system models locally and then combines models, reversing the sequence to preserve privacy while achieving predictive capability.
Solution Approach 2:
The patent segments the model building process into independent local training phases and a separate model combination phase. Each private data source independently trains its own statistical model, maintaining data locality and privacy. The segmentation allows predictive capability to be built through model ensembling without requiring centralized data access.
Data Source
AI summary
A distributed, online machine learning system is presented. Contemplated systems include many private data servers, each having local private data. Researchers can request that relevant private data servers train implementations of machine learning algorithms on their local private data without requiring de-identification of the private data or without exposing the private data to unauthorized computing systems. The private data servers also generate synthetic or proxy data according to the data distributions of the actual data. The servers then use the proxy data to train proxy models. When the proxy models are sufficiently similar to the trained actual models, the proxy data, proxy model parameters, or other learned knowledge can be transmitted to one or more non-private computing devices. The learned knowledge from many private data servers can then be aggregated into one or more trained global models without exposing private data.