AutoWordBug

Deep neural network based sentence classification attack.

Docker

We offer a docker image with all dependencies are installed. If you do not prefer using docker please skip this section and see the instructions from the next section (Requirements). You can build an image, called autowordbug, with the following command.

$ docker build . --tag autowordbug

Then, you can create a docker container with the created image with the following command.

$ docker run --rm -it autowordbug

Inside the container, we offer 3 system-wide programs, which are prepare, train, and evaluate. You can see how to use them by passing --help argument as follows:

$ prepare --help
$ train --help
$ evaluate --help

You also can get some hints at Training and Evaluation sections in this document.

Requirements

Following libraries are required. Install them using your pip.

• numpy
• pytorch

NOTE: Before Start

Before start, all scripts (.py files at the root directory) offer --help option. Please, see the usage message for more details which not explained in this document.

Training

You can train AutoWordBug with the following steps. First, you need to prepare the datasets. The following command will create train.pkl, val.pkl, and test.pkl at experiments folder.

$ python prepare.py

Then, you can train an AutoWordBug with the following command. If you want to utilize your GPU, type in --cuda option additionaly.

$ python train.py --train experiments/train.pkl --val experiments/val.pkl
# Training will take a while.

Now, you can find trained model at the experiments folder

Evaluation

With trained model, you can evaluate the performance of AutoWordBug.

$ python evaluate.py --model experiments/AutoWordBug_best.pt --test experiments/test.pkl
# Generation result of each sentences will be print out.

Maintainers

• Dohyun Kim
• Guhnoo Yun
• Jihyun Kim
• Seokhyun Lee