XONN: XNOR-based oblivious deep neural network inference

XONN is an oblivious inference framework based on Yao's garbled circuits, replacing matrix multiplication with XNOR operations that are nearly zero-cost in GC, and does not leak client inputs and results to the server. Its performance improves by up to 7 times over Gazelle, 93 times over MiniONN, and 37 times over SecureML. It also completes oblivious inference for the first time on the Fitnet architecture with up to 21 layers, and validates privacy-preserving medical diagnosis on four datasets.