TuringPass is a captcha solver detection system that determines whether a captcha is solved by a human or a bot. By analyzing behavioral biometrics and leveraging deep learning, TuringPass aims to enhance CAPTCHA security through anomaly detection and behavioral analysis — going beyond just checking if the solution is correct.

Traditional CAPTCHA systems are no longer reliable. Bots powered by advanced ML models can now solve text-based CAPTCHAs with high accuracy, making correctness alone an insufficient measure of human verification.

🔒 A better approach should:

• Validate the solution, and
• Analyze how the user arrives at it.

💡 This project explores both the attack vector (bot solving captchas) and the defense mechanism (detecting automated solvers using behavioral biometrics and anomaly detection).

• Captures user interactions while solving captchas using a Flask-based web interface.
• Uses an LSTM Autoencoder trained on human behavior to flag bot-like actions as anomalies.
• Simulates bot behavior via:
  • Random mouse movements
  • Timed keystrokes
  • OCR-based captcha solving using a CNN model
• Frontend is built using Tailwind CSS for a clean and responsive UI.

TuringPass/
│
├── backend/                  # Backend code
│   ├── app.py            # Contains endpoints for OCR inference and human/bot autoencoder inference
│   ├── captcha_model.h5  # OCR model 
│   ├── lstm_autoencoder_model.h5   # Trained LSTM Autoencoder
│   └── scaler.save          
│
├── models/                          # Model Training code
│   ├── captcha/                    # CAPTCHA solving module
│   │   ├── README.md
│   │   ├── human_captcha_solve_data.csv
│   │   ├── inference.ipynb
│   │   ├── ocr-model.ipynb
│   │   └── ocr_model.h5
│   │
│   ├── human-bot/                  # Human vs Bot anomaly detection module
│   │   ├── README.md
│   │   ├── anomaly.ipynb
│   │   ├── ci_main_cleaned.csv
│   │   ├── lstm_autoencoder_model.h5
│   │   ├── preprocessing.ipynb
│   │   └── scaler.save
│
├── static/
│   └── dataset/                         # CAPTCHA dataset images
│
├── templates/                           # HTML templates for frontend rendering
│
├── app.py                               # Flask file integrating human data collection and bot detection
├── captcha_interactions.csv             # Human interaction data collected 
├── captcha_solver_bot.js                # Advanced bot
├── captcha_solver_simple.js             # Script for simple CAPTCHA solving bot
├── demo.mp4                             # Demo video showcasing the project
├── requirements.txt                     # Python dependencies
└── README.md                         

To build a robust model for detecting non-human behavior, we collected 2,000+ CAPTCHA solving sessions from 20+ real participants, capturing fine-grained behavioral data during each interaction.

We recorded:

• 🖱 Mouse Coordinates
  Tracked continuously to analyze movement smoothness, path variance, and reaction time.
• ⌨️ Keystroke Events
  Logged every key press/release along with timestamps, including usage of backspace to detect natural corrections.
• ⏱ Timestamps
  Captured time per character, total session time, and inter-event delays to model decision-making pace.
• 🔡 Final Typed Response
  Compared with actual CAPTCHA values to validate correctness and analyze error tendencies.

These features were critical for:

• Training an LSTM Autoencoder on genuine human behavior patterns.
• Setting anomaly detection thresholds using reconstruction error distributions.
• Simulating and distinguishing realistic vs bot-like interactions.

With this multi-dimensional dataset, we created a behavioral fingerprint for every session—enabling precise and interpretable detection of automated activity.

• Tracks mouse coordinates, keypresses, and timing during captcha solving.
• Trains an LSTM Autoencoder on normal (human) behavior.
• At inference time, calculates reconstruction error — high error implies anomalous (bot) behavior.

• Programmatically simulates user actions with:
  • Noisy/randomized mouse movement paths
  • Fixed or abnormal timing for typing
• Uses a CNN-based OCR model to read captcha images and solve them.

• Custom CNN model trained to recognize alphanumeric characters from captchas.
• Dataset used: Kaggle - Captcha Dataset

git clone https://github.com/yourusername/TuringPass.git
cd TuringPass

pip install -r requirements.txt

cd app
python app.py

cd backend
python app.py

ML Models Used:

• 🧠 LSTM Autoencoder — Anomaly detection on behavioral data
• 🔎 CNN — Captcha OCR (Optical Character Recognition)

This project is licensed under the MIT License.
See the LICENSE file for more details.

• The captcha OCR model is trained using the open-source dataset: Kaggle - Captcha Dataset
• Inspired by real-world security challenges in distinguishing between bots and humans in critical applications.