Homomorphic Encryption at Scale
Overview
Our project explores advancements in Fully Homomorphic Encryption (FHE), a cutting-edge encryption method that allows for computations on encrypted data without needing to decrypt it first. This innovative approach to data privacy and security enables the performance machine learning operations on ciphertexts, ensuring that sensitive information remains secure even during processing.
One of the primary challenges with FHE is its computational inefficiency, significant space requirements, and the introduction of noise that complicates machine learning analyses on encrypted data. Our research team has focused on addressing these issues by implementing bootstrapping techniques to reduce noise and incorporating parallel processing to enhance function optimization.
In our specific application to linear regression models, we managed to increase computational precision and maintain data privacy without key dependency. By refining the dataset and employing bootstrapping, we achieved an increase in precision from 43 bits to 59 bits, and through parallel processing, we significantly reduced the computational time required for model training operations. Our approach also ensures that the precision of the linear regression model with homomorphic encryption is comparable to that of its unencrypted counterpart, albeit with careful attention to training iteration limits to avoid overfitting.
The outcome of our work is a python library that is open source and accessible for public use. This library leverages CPU power for parallel processing, thereby enhancing the speed and efficiency of computations on encrypted data. Our work represents a significant step towards making FHE more practical and accessible for real-world applications, particularly in fields like medical analysis where data privacy and security are crucial.
Community Benefit
Our project aims to contribute to a more private and efficient way of handling and analyzing encrypted data. This opens the door for companies and organizations to leverage powerful machine learning algorithms while ensuring the utmost confidentiality of their data. This has benefits in sectors like healthcare, finance, and any field where sensitive data handling is crucial. The use of CPU parallel processing to mitigate the computational drawbacks of FHE allows for handling larger datasets and more complex computations.
Team Members
Sponsored By
- AVAIntell-Marc Assailan