Adversarial Reprogramming

Introduction

This repo is a pytorch implementation of the paper. We aim to reprogram the pretrained imagenet models for MNIST classification. Following the code, you can easily add more datasets and other pretrained imagenet models for more experiments.

• The main_raw.py is the rough version from scratch.
• The main_single.py is the old version that is only able to train on a single gpu card.
• The main.py is the final version that supports multi-gpu training.

Requirements

• Python 2.7 or 3.x
• Pytorch 0.3
• tqdm

Training

1. Download pretrained imagenet models.

   ./download.py
   

2. The MNIST dataset will be automatically downloaded after running the scripts. Then, the structure of this repo should be as follows.

   ├── datasets
   │   └── mnist
   │       ├── processed
   │       │   ├── test.pt
   │       │   └── training.pt
   │       └── raw
   │           ├── t10k-images-idx3-ubyte.gz
   │           ├── t10k-labels-idx1-ubyte.gz
   │           ├── train-images-idx3-ubyte.gz
   │           └── train-labels-idx1-ubyte.gz
   ├── models
   │   ├── resnet50-19c8e357.pth
   │   └── vgg16-397923af.pth
   ├── train_log
   │   └── W_001.pt
   ├── README.md
   ├── download.py
   ├── main.py
   ├── ....
   
   

3. So directly run the following scripts to train the model.

   python main.py -m train -g 0 1
   

   We provide three three modes, train, validate and test. The argment -g indicates the gpu ids. Add -r 3 if you want to continue training your model from the third epoch.

Validate

Simply running the following command.

python main.py -m validate -g 0 -r 30

Results

The performance (test accuracy) of ImageNet model reprogrammed on other tasks. More experiments are welcomed by pull requests.

ImageNet Model	MNIST
Resnet50	0.94
VGG16	0.94