Privacy-Preserving Neural Network Model Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing privacy-preserving neural network models are inefficient and limited in practical applications due to slow homomorphic computations on encrypted data, especially for non-linear functions, and are not optimized for neural networks.
Innovation Solution
A method and system for building a privacy-preserving neural network model that involves performing initial operations on a non-private neural network in plaintext, encrypting the output, and then performing subsequent operations homomorphically using a private neural network, with learnable parameters trained during these operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If homomorphic encryption is used to perform computations on encrypted data, then privacy is protected, but computation speed deteriorates significantly
Solution Approach 1:
The patent segments the neural network into two distinct parts: a non-private neural network for initial plaintext processing and a private neural network for encrypted homomorphic processing. This segmentation allows the system to leverage the speed of plaintext computation for initial operations while reserving privacy-protected homomorphic computation for sensitive stages, thereby resolving the contradiction between privacy protection and computation speed.
Solution Approach 2:
The patent applies preliminary action by performing all non-private neural network operations on plaintext data before encryption. The non-private neural network processes input data in plaintext format to generate intermediate results, which are then encrypted. This preliminary plaintext processing avoids the need for slow homomorphic operations on input data, thus improving overall computation speed while maintaining privacy where necessary.
2Reliability
If homomorphic computations are performed on encrypted data, then privacy is maintained, but computation accuracy deteriorates due to noise accumulation
Solution Approach 1:
The patent divides the computation process into two segments: plaintext processing in the non-private neural network and encrypted processing in the private neural network. By performing critical feature extraction and initial transformations in plaintext, the system minimizes the number of homomorphic operations required, thereby reducing noise accumulation and improving prediction accuracy while maintaining privacy protection in the encrypted segment.
Solution Approach 2:
The patent introduces an intermediary mechanism through the encryption layer that connects the non-private and private neural networks. The encryption/decryption interface acts as a mediator that allows accurate plaintext results to be transferred to the encrypted domain, where further processing occurs with minimal noise introduction. This intermediary structure helps maintain accuracy across the privacy boundary.
3Reliability
If existing homomorphic encryption schemes are used for privacy-preserving neural networks, then privacy is protected, but practical applicability deteriorates due to poor performance
Solution Approach 1:
The patent segments the neural network architecture and processing pipeline to create a hybrid system that combines non-private and private components. This segmentation enables the system to be adapted to various practical applications by adjusting the proportion and configuration of private versus non-private layers, making privacy-preserving neural networks versatile enough for real-world deployment while maintaining strong privacy protection.
Solution Approach 2:
The patent employs parameter changes by optimizing the configuration of the hybrid neural network, including the number of layers in each segment, the encryption parameters, and the data transformation methods. By carefully tuning these parameters, the system achieves a balance between privacy protection strength and computational efficiency, thereby improving practical applicability across different application scenarios.
Data Source
AI summary
A method of building a privacy-preserving neural network model is provided. The method includes: performing first neural network operations using a non-private neural network of the privacy-preserving neural network model based on first input data in plaintext to produce first output data in plaintext, whereby the non-private neural network is pre-trained and learnable parameters of the non-private neural network are fixed while performing the first neural network operations; encrypting the first output data from the non-private neural network using a homomorphic encryption method to produce first encrypted data; and performing second neural network operations homomorphically using a private neural network of the privacy-preserving neural network model based on the first encrypted data to produce second encrypted data, whereby learnable parameters of the private neural network are trained while performing the second neural network operations. There is also provided a corresponding system for building a privacy-preserving neural network model, as well as method and system for performing privacy-preserving prediction using the privacy-preserving neural network model built.