Block-wise Scrambled Image Recognition Using Adaptation Network

This repository contains a Pytorch implementation of Proposed Adaptation Network in our AAAI WS 2020 paper:

Koki Madono, Masayuki Tanaka, Masaki Onishi, and Tetsuji Ogawa. Block-wise Scrambled Image Recognition Using Adaptation Network. In AAAI WS, 2020

Adaptation Network is described in Section "Adaptation Network for Block-WiseScrambled Image Recognition"

This network can be used for cloud based machine learning in visual information hiding setting.

Scrambling method is described in this figure.

Getting Started

We have tested our method on cifar10/100 Dataset (The dataset can be download via pytorch code)

Prerequisites

python
pytorch
numpy
scikit-image

Installing

1. Clone this repository: git clone https://github.com/MADONOKOUKI/aaai_ws.git

2. Install Pytorch.

3. pip install -r src/requirements.txt

Training

python {args.1}_{args.2}_{args.3} --e=305 --tensorboard_name * --training_model_name *.t7 --json_file_name *.json

args.1 : default(no adaptation network) / tanaka / proposed

args.2 : dataset(cifar10 / cifar100)

args.3 : encryption method(LE,ELE,EtC) or no encryption(plain)

We use shakedrop classifier as the backbone network with adaptation network.

Results

see mypaper

Result highly change depend on hyper parameter on matrix and total variation norm.

Citation

If you find this code useful for your research, please consider citing our paper:

@Inproceedings{madono2020,
  Title           = {Block-wise Scrambled Image Recognition Using Adaptation Network},
  Author         = {Koki Madono, Masayuki Tanaka, Masaki Onishi, and Tetsuji Ogawa},
  Booktitle      = {AAAI WS},
  Year           = {2020}
}

Reference codes

• mastnk/ICCE-TW2018
• owruby/shake-drop_pytorch
• utkuozbulak/pytorch-cnn-visualizations

License

MIT License. Please see the LICENSE file for details.