Federated Learning Model Integration for Confidential Data Protection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing federated learning techniques can potentially reveal confidential learning data by reverse engineering the global model, compromising data security.
Innovation Solution
A learning system that integrates a first machine learning model with a local model generated using confidential data and a global model trained on different data, generating a new global model without exposing the confidential data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If federated learning is used to train local models using data sets owned by respective organizations, then data confidentiality is maintained, but the organization can still infer learning data by reverse engineering the global model
Solution Approach 1:
The patent segments the data set into two distinct parts: first data (highly confidential) and second data (less confidential). The local model is trained only on the second data, while the first data is used separately to create a first machine learning model. This segmentation prevents the global model from containing highly confidential data that could be inferred through reverse engineering, while still allowing the organization to benefit from federated learning.
Solution Approach 2:
The patent introduces an intermediary approach by creating a local model that serves as a mediator between the global model and the highly confidential first data. The local model is trained on second data and then integrated with the first machine learning model (trained on first data) to create a second machine learning model. This intermediary structure allows the system to maintain data confidentiality while achieving accurate machine learning.
2Measurement precision
If the global model is used to learn confidential data, then the accuracy of the machine learning model is improved, but the confidential data may be leaked through the global model
Solution Approach 1:
The patent segments the data into first data (highly confidential) and second data (less confidential), and trains different models on different data segments. The local model is trained only on second data, preventing confidential data from being directly used in the global model. The first machine learning model trained on first data is then integrated with the local model to achieve high accuracy without leaking confidential information.
Solution Approach 2:
The patent extracts the highly confidential first data from the training process of the global model. Instead of training the global model on all data including highly confidential data, the patent extracts and separates the first data, using it only to create the first machine learning model which is then integrated with the local model. This extraction prevents data leakage while maintaining model accuracy.
3Measurement precision
If all data items are used to train the global model, then the model accuracy is maximized, but the risk of data leakage increases
Solution Approach 1:
The patent segments the data set into first data (highly confidential) and second data (less confidential). The global model is trained only on second data, while the first data is used to create the first machine learning model. This segmentation allows the system to maximize model accuracy by using all data for training while minimizing data leakage risk by preventing highly confidential data from being used in the global model.
Solution Approach 2:
The patent applies local quality by treating different data items differently based on their confidentiality level. The first data (highly confidential) is processed differently from the second data (less confidential). The local model is trained only on second data, while the first data is used separately. This local quality approach allows the system to maximize accuracy by utilizing all data while managing leakage risk through differential processing.
Data Source
AI summary
A learning system includes: a first learning unit configured to cause a global model generated by federated learning to learn first data included in the data set, to thereby generate a local model; a second learning unit configured to learn a first machine learning model by machine learning that uses second data among the data items included in the data set, the second data being different from the first data; an integration unit configured to integrate the local model or the global model with the first machine learning model; and a generation unit configured to generate a new global model using the local model.