Homomorphic Encryption Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning technologies face challenges in training deep neural networks on encrypted data due to computational complexity and privacy concerns, where sensitive data cannot be shared, and existing solutions are either too slow or do not address the training phase effectively.
Innovation Solution
The implementation of Fully Homomorphic Encryption (FHE) based Network for Inferencing and Training (FHE-NIT) system, which uses the open-source FHE toolkit HElib to perform Stochastic Gradient Descent-based training of neural networks on encrypted data, allowing for inference operations without revealing sensitive information and enabling secure model training and inference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data is encrypted to preserve privacy, then data security is improved, but machine learning training becomes computationally complex and slower
Solution Approach 1:
The patent uses homomorphic encryption as an intermediary that allows computations to be performed on encrypted data without decryption. The encryption scheme acts as a mediator between the need for data security and the requirement for efficient training, enabling the cloud server to process encrypted data as if it were plaintext through specially designed homomorphic operations.
Solution Approach 2:
The patent transforms the machine learning training process to work with encrypted data by changing the computational parameters and operations. Instead of standard plaintext operations, the system uses homomorphic encryption parameters and corresponding encrypted operations, allowing training to proceed on encrypted data while maintaining security.
2Measurement precision
If data is shared for model training, then model accuracy is improved, but data privacy is compromised
Solution Approach 1:
Homomorphic encryption serves as an intermediary that enables the cloud server to access and process the end user's data for training purposes without actually seeing or exposing the plaintext data. The encryption layer allows the server to perform necessary computations while the data remains confidential, thus achieving both model training and privacy preservation.
3Reliability
If encrypted data is used for training, then data privacy is preserved, but computational complexity increases
Solution Approach 1:
The patent extracts and separates the computational complexity of homomorphic encryption operations from the standard machine learning training process. By using specialized homomorphic encryption schemes and optimized operations, the system isolates the complex cryptographic computations from the general-purpose training algorithms, allowing each to be optimized independently.
4Power
If cloud computing is used for training, then computational power is improved, but data security concerns increase
Solution Approach 1:
The patent employs homomorphic encryption as an intermediary layer between the end user's data and the cloud computing infrastructure. This allows the cloud server to leverage its substantial computational power for training deep neural networks while the encryption scheme ensures that the server never accesses or exposes the actual data, thus resolving the trust and security concerns associated with cloud-based training.
Data Source
AI summary
A computer system receives a set of data encrypted by a homomorphic encryption transformation. The computer system performs machine learning operations using the encrypted set of data. The machine learning operations build, using homomorphic operations, a trained model of the data having a mapping between the encrypted data and output of the trained model. The model is stored for use for performing inferencing of other encrypted data to determine a corresponding output of the trained model. The computer system may perform inferencing of the other encrypted data at least by accessing the stored trained model and predicting by using the trained model a label in an encrypted format that corresponds to the other encrypted data. The computer system may send the label toward the client for the client to decrypt the label.