Aggregate Model Using Feature Contribution Values for Privacy-Preserving Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data protection techniques in multi-party environments, such as vertical federated learning, lack a reliable method to combine models without sharing source data, compromising data privacy and resulting in inferior collaboration outcomes among companies.
Innovation Solution
The use of feature contribution values, like SHAP values, from multiple federated machine learning models as inputs to an aggregate machine learning model allows for global predictions without exposing raw data, maintaining privacy and security while enabling collaboration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If source data is shared among parties for model combination, then model collaboration effectiveness is improved, but data privacy is compromised
Solution Approach 1:
The patent extracts only the necessary predictive information (feature contribution values) from the source data, separating it from the sensitive raw data. Each party's model generates SHAP values that capture the contribution of each feature to predictions, and these extracted values are shared instead of the original data, enabling collaboration while preserving privacy.
Solution Approach 2:
The patent introduces feature contribution values (SHAP values) as an intermediary that mediates between the need for data sharing and privacy protection. These intermediary values serve as a proxy that allows models to be combined and predictions to be made without directly exposing the sensitive source data between parties.
2Object-affected harmful factors
If feature contribution values are shared instead of source data, then data privacy is protected, but model combination reliability is reduced
Solution Approach 1:
The patent transforms the data representation from raw source data to feature contribution values (SHAP values), changing the parameter space. This transformation maintains the essential predictive information while removing sensitive data, allowing reliable model combination through mathematical aggregation of the transformed parameters.
Solution Approach 2:
The patent creates a composite approach by combining multiple parties' feature contribution values into an aggregate model. The composite model integrates the SHAP values from different parties' models, maintaining reliability through the aggregated information while preserving the privacy benefits of not sharing raw data.
3Measurement precision
If multiple federated models are aggregated, then prediction accuracy is improved, but system complexity increases
Solution Approach 1:
The patent segments the complex task of multi-party collaboration into manageable components: each party independently trains their own model, generates feature contribution values, and these segmented results are then aggregated. This segmentation reduces system complexity by allowing independent parallel processing while achieving improved prediction accuracy through combination.
Solution Approach 2:
The patent uses feature contribution values as a copy or representation of the underlying model behavior. Instead of sharing or directly combining complex models, the system creates simplified copies in the form of SHAP values that capture the essential predictive patterns, reducing complexity while maintaining accuracy.
Data Source
AI summary
Techniques are provided for multi-party prediction using feature contribution values. One method comprises obtaining a first set of feature contribution values associated with respective ones of a plurality of machine learning models, wherein each machine learning model is trained using training data of a different party and each feature contribution value indicates a contribution by a corresponding feature to a prediction generated by the associated machine learning model; training an aggregate machine learning model using the obtained first sets of feature contribution values; receiving a second set of feature contribution values generated by applying data of at least one party to at least one machine learning model; and applying the second set of feature contribution values to the trained aggregate machine learning model to obtain a global prediction. Each feature contribution value may correspond to a masked feature, and the feature contribution values may not expose the source data of one party to another party.