ACAPTCHA

ACHAPTCHA is composed of two components: (1) an adversarial text captcha generation method, and (2) an adversarial text captcha attack method.

Prerequisites

• Install necessary packages from requirements.txt
• Download crnn.rar from here and unzip them into crnn

Getting Started

1. Generate host captchas including different security features

Modify the python file into gen_v*.py, where * represents the version number(i.e., 1~6).

cd design
python gen_v1.py

2. Add adversarial example on HCAPTCHA

Firstly, you need to train attack models. Data will be stored in lmdb.

cd crnn
python tool/create_dataset.py --out lmdb/version/v1_1/train --folder ../datasets/version/v1_1/train
python tool/create_dataset.py --out lmdb/version/v1_1/val --folder ../datasets/version/v1_1/val
python train.py --trainroot lmdb/version/v1_1/train --valroot lmdb/version/v1_1/val

Then, you can generate adversarial examples using different adversarial example algorithms (fgsm.py, i-fgsm.py, or deepfool.py).

python fgsm.py

3. Eliminate perturbation of ACAPTCHA through image processing techniques

We integrate 9 image processing techniques (Pix2Pix, total variation loss, three data compression algorithms, and four image filters) to remove adversarial perturbation in ACAPTCHAs.

python filter.py

4. Solve the processed captchas using attack models

python test.py