Secure Outsourced Matrix Computation and Application to Neural Networks.

Area of Science:

• Cryptography
• Privacy-Preserving Machine Learning
• Computational Security

Background:

• Homomorphic Encryption (HE) offers privacy for outsourced computations but lacks efficient matrix operations for modern machine learning.
• Secure Multi-Party Computation (MPC) is an alternative but has higher communication costs and is less suitable for non-interactive tasks.

Purpose of the Study:

• To develop a practical solution for homomorphic encryption of matrices and enable secure arithmetic operations on them.
• To enhance the applicability of HE in machine learning frameworks by addressing limitations in matrix computation.
• To introduce a novel framework (E2DM) for secure prediction using both encrypted data and encrypted models.

Main Methods:

• A novel matrix encoding method for homomorphic encryption.
• An efficient evaluation strategy for basic matrix operations (addition, multiplication, transposition).
• Techniques for encrypting multiple matrices in a single ciphertext for improved performance.

Main Results:

• Demonstrated practical performance for encrypted matrix multiplication (64x64 matrices in 9.21s) and transposition (64x64 matrices in 2.56s).
• Developed a generic solution applicable to most existing HE schemes.
• Achieved secure prediction on the MNIST dataset using convolutional neural networks (CNNs) with encrypted data and models, processing 64 images in 28.59s.

Conclusions:

• The proposed HE-based matrix computation mechanism is practical and efficient, overcoming previous limitations in secure machine learning.
• The E2DM framework enables secure prediction with both encrypted data and models, a novel advancement in the field.
• This work significantly advances the potential of HE for real-world privacy-preserving machine learning applications.