Server System Combining ML Models via Metadata for Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning model inference systems face challenges in maintaining user privacy, particularly when combining user-created models, as they often require access to sensitive training data.
Innovation Solution
A system comprising mobile computing devices, a server system, a lockbox node, and a deployment node that allows users to create and refine machine learning models on their devices without sharing the underlying training data, with the server system grouping and combining models based on metadata to perform inference tasks while maintaining privacy through secure data storage and controlled access.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If user training data is shared to combine and improve machine learning models, then model performance is improved, but user privacy is compromised
Solution Approach 1:
The patent extracts only the essential model parameters (weights, biases, architecture) from the complete training data, separating the useful information needed for model combination from the sensitive personal data. This allows model improvement without exposing user privacy.
Solution Approach 2:
Instead of sharing actual training data, the system creates and shares copies of model parameters that capture the learned patterns. These parameter copies enable model combination while the original sensitive training data remains private on user devices.
2Object-affected harmful factors
If only model parameters are shared instead of training data, then user privacy is protected, but model combination accuracy may be reduced
Solution Approach 1:
The patent creates a universal model parameter format that can represent different model types and architectures in a standardized way. This universal representation enables accurate combination of diverse models while maintaining privacy protection.
Solution Approach 2:
The system transforms model parameters into a standardized representation format that preserves the essential information needed for accurate model combination. By changing the parameter representation rather than sharing raw data, both accuracy and privacy are maintained.
3Manufacturing precision
If a centralized system collects all training data for model creation, then model accuracy is improved, but system complexity and security risks increase
Solution Approach 1:
The patent segments the model creation process into distributed independent training phases on user devices, followed by a centralized parameter aggregation phase. This segmentation reduces system complexity by allowing parallel independent training while maintaining accuracy through coordinated parameter combination.
Solution Approach 2:
The system introduces model parameters as an intermediary between local training data and centralized model combination. This intermediary enables accurate model aggregation without requiring the centralized system to handle or process sensitive raw training data, reducing security risks and complexity.
Data Source
AI summary
A server system is described. The server system includes at least one processor and at least one memory storing a plurality of instructions that, when executed by the at least one processor, causes the at least one processor to receive a plurality of machine learning (ML) models and ML model data from the plurality of mobile computing devices, where the ML model data includes metadata, and the ML model data excludes media. At least one group of the plurality of ML models is selected based on the metadata, each ML model of the at least one group tested, and a plurality of metrics is determined based on the ML models that were tested. Further, a set of ML models of the ML models that were tested is selected, and at least two ML models of the set of ML models are combined based on at least one parameter.