cuSCNN: A Secure and Batch-Processing Framework for Privacy-Preserving Convolutional Neural Network Prediction on...

Area of Science:

• Computer Science
• Cryptography
• Machine Learning

Background:

• Privacy-preserving deep learning as a service is gaining traction for secure cloud-based neural network predictions.
• Existing solutions using BGV-based homomorphic encryption on CPUs are inefficient for large datasets due to complex computations.
• Matrix multiplication is a critical, time-consuming operation in secure linear layers.

Purpose of the Study:

• To develop an efficient and practical framework for privacy-preserving deep learning predictions on GPU.
• To address the performance limitations of CPU-based homomorphic encryption schemes in deep learning.

Main Methods:

• Proposed cuSCNN framework utilizing GSW-based homomorphic matrix encryption for CNN linear layers.
• Implemented a hybrid optimization approach with CUDA (Compute Unified Device Architecture) for enhanced GPU parallelism and memory access.
• Focused on optimizing the atomic matrix multiplication operation for efficiency.

Main Results:

• cuSCNN demonstrates superior performance in runtime and power consumption compared to existing frameworks.
• The GSW-based encryption efficiently secures matrix computations, avoiding BGV-scheme complexities.
• CUDA optimization significantly boosts computation efficiency on the GPU platform.

Conclusions:

• cuSCNN offers a secure, efficient, and practical solution for privacy-preserving deep learning on GPUs.
• The framework effectively reduces the time cost of secure linear layers in CNNs.
• This approach paves the way for more scalable and energy-efficient cloud-based AI services.