Breaking Wikipedia's Captcha

---

Breaking 2020 Wikipedia's captcha using image processing and deep learning techniques. Before publishing the project, I sent an email to Wikipedia to make them aware of the vulnerability. I'm not responsible for how this code is used.

👓 Demo

🎢 Process

I try to demonstrate how simple image processing techniques and a shallow model like LeNet is enough to break some captcha systems. The project aims to leverage traditional computer vision techniques to train a neural network capable of distinguishing between characters. Before publishing the project, I sent an email to Wikipedia to make them aware of the vulnerability. I’m not responsible for how this code is used.

1. Collect some random captcha images from Wikipedia

After so many failed attempts to log-in, you are displayed with a captcha image. Each captcha image has 9 or 10 characters. I collected 230 images, so I would have at least 9x230=2070 character images.

2. Extract the characters from the images and label them

Once I had the images, I applied some traditional Computer Vision techniques:

• Binarization (OTSU and manual) to improve the quality of the characters
• Contours to find and extract the characters individually
• Masks to merge to characters together (the dot of the “i” and the vertical line)
• I also had to consider the characters that don’t have any space between them and are detected as a single one.
• All the characters are saved in the corresponding folder inside of the dataset directory after manually labeling them.

3. Train and fine-tune the neural network

I used the individual character images with size 28x28 to train a LeNet model. The model achieved 96% accuracy with 24 classes and only 20 epochs. I also used momentum to make it converge quicker.

4. Test it

With new captcha images, the individual characters are extracted, and then I use the already trained model to predict the text. You can see a working demo here.

📄 Results

I used the LeNet architecture, first introduced by LeCun et al. in their 1998 paper, Gradient-Based Learning Applied to Document Recognition. The authors' motivation for the implementation of LeNet was primarily for Optical Character Recognition (OCR). It's a simple model with only two convolutional layers. I achieved 96% accuracy and 0.15 loss both on the train and the test set.

🔧 Setup

git clone https://github.com/pauladj/ml-breaking-wikipedia-captcha.git
cd ml-breaking-wikipedia-captcha
pip install -r requirements.txt

🎈 Usage

You can use the already downloaded captchas in the downloads folder or you can download more images using:

python download_images.py --output <captcha_image_folder> -n <num_images_to_download>

To get the text of the new captcha images you just have to execute the next command:

python test_model.py --input <captcha_image_folder> --model output

⛏️ Built Using

• OpenCV - Computer vision library
• Keras - Deep learning library
• Streamlit - Deploy data apps

🎉 Acknowledgements

• Inspired by Adrian Rosebrock's book Deep Learning for Computer Vision with Python (Starter Bundle)