Edge Device ML Model Sharing via Incremental Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing architecture for training machine learning models at a central server raises privacy and security concerns, as user data is aggregated and potentially hacked, leading to data leaks and denial of benefits from ML models.
Innovation Solution
A processor-implemented method for sharing an incrementally trained machine learning model from one edge device to others in a peer-to-peer network, involving generating a unique persistent file format, encrypting it for authentication, and transmitting it to other edge devices for incremental training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data is sent from edge devices to a central server for ML model training, then the ML model can be trained using aggregated data, but user data privacy and security are compromised due to potential hacking and data leaks
Solution Approach 1:
The system segments the centralized ML training process into distributed edge-based training units. Each edge device independently trains local ML models using its own data, eliminating the need to aggregate sensitive user data at a central server. This segmentation maintains data security while preserving training effectiveness through local model development.
Solution Approach 2:
The patent introduces an intermediary mechanism where edge devices share anonymized training insights and model updates rather than raw user data. This intermediary approach allows collaborative ML model improvement across the network while preventing direct exposure of sensitive user information, thus maintaining both security and training productivity.
2Measurement precision
If user data is aggregated at a central server for ML training, then comprehensive models can be built, but the risk of data leaks and malicious hacking increases
Solution Approach 1:
The system extracts only the essential training signals and model parameters from edge devices, leaving the sensitive raw user data localized on each device. By taking out only the necessary information for model improvement while removing the harmful exposure of complete user datasets, the system achieves model accuracy without data leak risks.
Solution Approach 2:
Each edge device maintains local control over its data quality and training process, applying local quality standards to protect sensitive information. The system allows each device to contribute to model accuracy through locally processed information while maintaining different levels of data protection tailored to local security requirements.
3Loss of time
If ML models are trained periodically at a central server, then models can be updated, but the learning process is slow and not continuous
Solution Approach 1:
Edge devices perform preliminary ML model training continuously in the background before any central coordination is needed. This preliminary action allows models to be pre-trained locally using incoming data streams, eliminating waiting time and enabling immediate deployment of updated models without periodic centralized processing delays.
Solution Approach 2:
The system establishes continuous ML training operations at edge devices rather than periodic centralized training. Useful training actions continue uninterrupted as data arrives, with models being continuously refined using incoming information streams, thereby eliminating idle time between training cycles and maximizing learning speed.
4Ease of operation
If large ML models are repeatedly transmitted to the server, then model updates can be deployed, but network load increases significantly
Solution Approach 1:
Instead of transmitting complete large-scale ML models, the system transmits only partial updates containing the essential changes and parameter adjustments. This partial action approach delivers the necessary model deployment information while significantly reducing transmission volume and network bandwidth consumption.
Solution Approach 2:
The system creates lightweight copies of model update information rather than transmitting full model replicas. These simplified copies contain only the critical parameter changes needed for deployment, enabling easy model updates while minimizing network energy loss through reduced data transmission volumes.
Data Source
AI summary
A method for sharing an incrementally trained machine learning (ML) model from an edge device to other edge devices that are associated with a peer to peer network is provided. The method includes (i) receiving event data at an edge device from among one or more edge devices of the peer to peer network upon the occurrence of an event at the edge device, (ii) incrementally training a base version of a ML model at the edge device based on the received event data, (iii) generating, a unique persistent file format of the incrementally trained ML model at the edge device, (iv) optimizing a payload associated with the unique persistent file format to include one or more parameters with associated weights changing beyond a predetermined configurable threshold and (v) encapsulating the unique persistent file format with a unique metadata.