GAZELLE: A Low Latency Framework for Secure Neural Network Inference
summary
Describes Gazelle, a system that combines lattice-based homomorphic encryption (with SIMD, multiplication and ciphertext permutation primitives) and two-party techniques (garbled circuits) to perform private CNN inference. Presents optimized HE linear-algebra kernels for matrix-vector multiplication and convolution, plus protocols that switch encodings between HE and garbled circuits, and reports large online runtime speedups over MiniONN, Chameleon and CryptoNets on MNIST and CIFAR-10.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:32Motivation
- 4:16Secure Computation (2PC)
- 7:04Secure Neural Network Inference
- 8:57FHE vs 2PC vs PAHE+2PC
- 13:44The PAHE Abstraction
- 15:48Gazelle: Core Library
- 17:02Gazelle Core - Performance
- 19:42Convolution Layers
- 20:05SISO Convolution - Simple
- 21:58Gazelle Network - Performance
related talks
15mGALA: A High Performance Graph Neural Network Acceleration LAnguage and Compiler
28mGenerative AI + Proprietary Data
15mCocoon: A System Architecture for Differentially Private Training with Correlated Noises
35mEncrypted Computation: What if Decryption Wasn’t Needed?
19mSOTER: Guarding Black-box Inference for General Neural Networks at the Edge
56mChannel Gating for Cheaper and More Accurate Neural Nets
42mPrivacy Challenges in Machine Learning
14mKairox: Adaptive GPU-CPU Hybrid LLM Inference via Online Neuron Balancing