GAZELLE - A Low Latency Framework for Secure Neural Network Inference

Explore a groundbreaking framework for secure neural network inference in this USENIX Security '18 conference talk. Delve into the Gazelle system, which combines homomorphic encryption and two-party computation techniques to address privacy concerns in cloud-based machine learning. Learn about the Gazelle homomorphic encryption library, optimized homomorphic linear algebra kernels, and encryption switching protocols that enable efficient and private neural network predictions. Discover how Gazelle outperforms existing systems in online runtime and offers significant improvements over fully homomorphic approaches. Gain insights into secure computation, homomorphic encryption, and their applications in protecting both client input and server neural network privacy in prediction-as-a-service scenarios.

Intro
Motivation
Secure Computation (2PC)
Secure Neural Network Inference
FHE vs 2PC vs PAHE+2PC
The PAHE Abstraction
Gazelle: Core Library
Gazelle Core - Performance
Convolution Layers
SISO Convolution - Simple
Gazelle Network - Performance