Federated Learning Feature Space Expansion via Encrypted Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current lateral federation learning methods are limited to scenarios where both parties have labeled samples with the same feature dimensions, and cannot expand the feature space while ensuring data privacy and model security, which restricts the predictive ability of the federation model.
Innovation Solution
A model parameter training method based on federation learning that determines a feature intersection between two parties, trains a mapping model, predicts missing features using encryption, and calculates loss values to determine convergence, allowing for the expansion of the feature space and improvement of predictive ability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If federation learning is used to protect data privacy, then data privacy and model security are improved, but the feature space expansion capability deteriorates
Solution Approach 1:
The patent introduces a third party (third terminal) as an intermediary to facilitate feature space expansion. The third terminal receives encrypted mapping models from both parties, decrypts them, and uses them to predict missing features without accessing the original data. This intermediary approach enables feature space expansion while maintaining the privacy protection benefits of federation learning.
Solution Approach 2:
The patent transforms the feature space by introducing a mapping model dimension that connects the feature spaces of both parties through the intersection features. By creating a mapping from intersection features to non-intersection features, the system expands the effective feature space without directly combining the original feature spaces, thus maintaining privacy while enabling expansion.
2Measurement precision
If data is aggregated to improve predictive ability, then predictive ability is improved, but data privacy compliance deteriorates
Solution Approach 1:
The patent creates encrypted copies of the mapping models that can be shared between parties without exposing the original data. The third terminal receives and uses these encrypted copies to perform feature space expansion and predictive modeling, achieving improved predictive ability without aggregating or exposing the original data, thus maintaining privacy compliance.
Solution Approach 2:
The patent segments the feature space into intersection features and non-intersection features, and segments the mapping process into separate encrypted mapping models. This segmentation allows each party to contribute only their local mapping model to the third terminal, enabling collaborative predictive ability improvement without requiring data aggregation or compromising privacy.
3Ease of manufacture
If lateral federation learning is used with same feature dimensions, then model training is simplified, but feature space expansion capability deteriorates
Solution Approach 1:
The patent extends the lateral federation learning approach by introducing a mapping dimension that handles different feature dimensions. Instead of requiring identical feature spaces, the system creates a mapping model that transforms intersection features to non-intersection features, enabling feature space expansion while maintaining the collaborative training simplicity of federation learning.
Solution Approach 2:
The patent makes the mapping model universal by training it to handle multiple feature dimension combinations. The same mapping model structure can process different intersection and non-intersection feature sets, making the system adaptable to various feature space configurations while maintaining the ease of collaborative training through encrypted model sharing.
Data Source
AI summary
Disclosed are a model parameter training method and a terminal based on federation learning, and a medium. The method includes: determining a feature intersection of a first sample of the first terminal and a second sample of a second terminal, training the first sample based on the feature intersection to obtain a first mapping model, sending the first mapping model to the second terminal; receiving a second encryption mapping model sent by the second terminal, predicting a missing feature of the first sample of the first terminal according to the second encryption mapping model to obtain a first encryption supplementary sample; receiving a first encryption federation learning model parameter sent by a third terminal, training a federation learning model to be trained according to the first encryption federation learning model parameter, and calculating a first encryption loss value; and sending the first encryption loss value to the third terminal.