Homomorphic Encryption for Secure Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network models using homomorphic encryption only perform operations during inference and do not effectively utilize homomorphically encrypted messages for training, leading to exposure of training data when used across multiple devices.
Innovation Solution
An electronic apparatus and server apparatus system that homomorphically encrypts and processes training data and model features, allowing secure training of neural network models without exposing the data, using matrix multiplication calculations and communication protocols to manage training operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If homomorphic encryption is applied during inference process, then security of encrypted messages is improved, but training data exposure risk increases when used across multiple devices
Solution Approach 1:
The system segments the training process into multiple stages: local preprocessing of training data, secure transmission of encrypted data segments to the server, distributed training operations on encrypted segments, and aggregation of results. This segmentation allows homomorphic encryption to be applied selectively during inference while maintaining security during training through encrypted data segments.
Solution Approach 2:
The patent introduces an intermediary encryption layer that mediates between the training data and the neural network model. Homomorphically encrypted training data segments are transmitted through this intermediary layer to the server, which performs training operations without decrypting the data, thus preventing direct exposure while enabling collaborative training across multiple devices.
2Measurement precision
If training data is transmitted to server for model training, then model accuracy is improved, but data privacy is compromised
Solution Approach 1:
The system performs preliminary homomorphic encryption of training data segments before transmission to the server. This preliminary action ensures that data privacy is protected from the outset, while the encrypted segments still contain sufficient information for the server to perform accurate training operations and improve model accuracy.
Solution Approach 2:
The patent changes the state of training data from plaintext to homomorphically encrypted form, altering its parameters to enable secure transmission and processing. The encrypted segments maintain the mathematical properties necessary for training operations while eliminating privacy risks associated with plaintext transmission.
3Ease of operation
If decryption is performed for calculation operations, then computational flexibility is improved, but security against third-party attacks deteriorates
Solution Approach 1:
The system extracts and separates the decryption operation from the training process. Instead of decrypting data for calculations, the patent extracts only the necessary computational operations and performs them directly on homomorphically encrypted data segments at the server, maintaining security while achieving computational flexibility through homomorphic properties.
Solution Approach 2:
The patent substitutes the traditional mechanical decryption-calculation-encryption process with a homomorphic encryption-based system where calculations are performed directly on encrypted data. This substitution eliminates the need to expose plaintext during computation, maintaining security while providing the same computational flexibility through mathematical properties of homomorphic encryption.
Data Source
AI summary
Disclosed is an electronic apparatus. The electronic apparatus includes: a memory configured to store a training model trained using a plurality of training data; a communication device configured to perform communication with an external device; and a processor, in which the processor is configured to control the communication device to extract a feature of the training model, and homomorphically encrypt the extracted feature and transmit the homomorphically encrypted feature to a server apparatus, control the communication device to homomorphically encrypt the plurality of training data and transmit the homomorphically encrypted training data to the server apparatus, and when a homomorphically encrypted calculation result is received from the server apparatus, control the communication device to decrypt the received calculation result and transmit information on a training state corresponding to the decrypted value to the server apparatus.