Homomorphic Encryption Compatible Machine Learning Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models face challenges in being trained using encrypted data and outsourcing training to third parties without compromising sensitive internal parameters, as conventional methods are computationally intensive and require interactive multi-party interactions, which are inefficient and insecure.
Innovation Solution
Modifying machine learning models to be compatible with homomorphic encryption by replacing non-compatible functions with polynomial functions, allowing training and classification using encrypted data while keeping internal parameters private, and using a third party for training without exposing sensitive data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are trained using conventional methods, then training can be performed with standard functions, but the process is computationally intensive and cannot be securely outsourced to third parties
Solution Approach 1:
The patent transforms the machine learning model by replacing non-compatible functions (ReLU, Batch Normalization, Softmax) with polynomial approximations that are compatible with homomorphic encryption. This parameter change in the mathematical representation of functions enables encrypted computation while maintaining model functionality, allowing secure outsourcing to third parties without compromising internal parameter security.
Solution Approach 2:
The patent substitutes conventional training mechanisms with homomorphic encryption-based training. By replacing standard computational operations with cryptographically secure polynomial operations, the system enables training to be performed on encrypted data by third parties without exposing sensitive model parameters, thus improving security while maintaining training capability.
2Reliability
If non-compatible functions are replaced with polynomial functions for homomorphic encryption compatibility, then secure encrypted training becomes possible, but model accuracy may be reduced due to approximation errors
Solution Approach 1:
The patent applies polynomial approximation of limited degree to replace complex functions. By using sufficiently high-degree polynomials (e.g., degree 4 or higher), the approximation error is reduced to acceptable levels while maintaining compatibility with homomorphic encryption. This partial replacement strategy balances security requirements with model accuracy preservation.
Solution Approach 2:
The patent performs preprocessing steps including adding bias terms and adjusting polynomial coefficients before training to compensate for approximation errors. These preliminary adjustments help mitigate the impact of function replacement on model accuracy, ensuring that the encrypted training process produces results comparable to conventional training.
3Reliability
If interactive multi-party computation is used for secure training, then privacy can be protected, but the process becomes complex and requires coordination between multiple parties
Solution Approach 1:
The patent extracts the security requirement from the training process by using homomorphic encryption to enable secure computation without requiring interactive verification between multiple parties. The encryption scheme itself provides the security guarantee, eliminating the need for complex multi-party protocols and reducing system coordination complexity while maintaining privacy protection.
Data Source
AI summary
A computer-implemented method for training a machine learning model, the method comprising: obtaining a machine learning model comprising a plurality of computational layers, the layers being arranged such that outputs from one or more of the layers serve as inputs to other ones of the layers; identifying one or more of the layers as comprising one or more functions that are not compatible with a homomorphic encryption scheme; replacing the one or more functions with alternative functions, wherein the alternative functions are functions that are compatible with the homomorphic encryption scheme and which provide an approximation of the respective functions that they replace; and sending the model to a third party to train the model using a set of training data.