Privacy-Preserving Image Classification Using an Isotropic Network

This repository is the official implementation of Privacy-Preserving Image Classification Using an Isotropic Network.

The main idea of this paper is to show EtC encrypted images can be learned by modern networks like ViT.

• EtC classification on ConvMixer is based on ConvMixer.
• EtC classification on ViT is based on ViT-pytorch.

Please follow the above repositories for setup.

Applying EtC on each image batch is

from encryption import EtC
etc = EtC(14, 16) # the image size is 224 and EtC uses block size 16 => 16 * 14 = 224

# then in your training loop
for X, y in dataloader:
    X = etc(X)