Learn about CryptGPU, a system for fast privacy-preserving machine learning on GPUs, in this 15-minute IEEE conference talk. Explore the challenges of privacy-preserving machine learning, including the threat model and image classification. Discover how CryptGPU addresses the performance gap and scalability issues through GPU acceleration. Gain insights into the system overview, design, and GPU-friendly protocol design. Understand the implementation of floating-point arithmetic and private training techniques. Conclude with a comprehensive understanding of how CryptGPU enhances privacy and performance in machine learning applications.
CryptGPU: Fast Privacy-Preserving Machine Learning on the GPU
Overview
Syllabus
Introduction
PrivacyPreserving Machine Learning
Threat Model
Image Classification
Performance Gap
Scalability
GPU acceleration
System overview
System design
Floating Point Arithmetic
GPUFriendly Protocol Design
Private Training
Conclusion
Taught by
IEEE Symposium on Security and Privacy
