scramblekit

Image scrambling for privacy-preserving deep learning: the block-wise scrambling schemes LE, ELE, EtC and pixel-based encryption, the adaptation networks that classify scrambled images, scrambling parameter generation, the SIA-GAN inversion attack and ScrambleMix — the official library of four papers by Koki Madono et al., with the keys and conventions of the original research code (checked bit for bit against it).

pip install scramblekit            # scrambling, PSNR/SSIM, SPG with SSIM - NumPy + Pillow only
pip install "scramblekit[all]"     # + PyTorch/torchvision (networks, SIA-GAN, ScrambleMix) + LPIPS
scramblekit demo --out demo/       # scramble a bundled sample image with every scheme (CPU, ~1 s, no downloads)

import numpy as np
import scramblekit
from scramblekit.scramble import ELE

img = scramblekit.sample_image()                    # 128x128 RGB uint8 (public-domain photo)
ele = ELE(key=0, image_size=128)                    # extended learnable encryption, 4x4 blocks, seed 0
scrambled = ele(img)                                # also PIL images and (N, 3, H, W) torch tensors
assert np.array_equal(ele.inverse(scrambled), img)

from scramblekit.models import build_model          # needs scramblekit[torch]
net = build_model("shakepyramidnet", num_classes=10, adaptation="ele")  # ELE-AdaptNet + ShakeDrop PyramidNet-110

Command line: scramblekit demo | train | spg | attack | cite (--dataset fake for smoke runs without downloads). Documentation, the Colab notebook, the list of paper-vs-code differences and the links to the paper repositories: https://github.com/MADONOKOUKI/scramblekit

If scramblekit helps your research, please cite the paper(s) of the modules you use: scramblekit cite (or scramblekit.cite(name)) prints the BibTeX entries.