README.md

metadata

language:
  - en
license: mit
size_categories:
  - 1K<n<10K
task_categories:
  - image-classification
  - feature-extraction
tags:
  - captcha
  - security
  - ai-resistance
  - adversarial-robustness
  - computer-vision
pretty_name: DeepCaptcha Dataset
configs:
  - config_name: default
    data_files:
      - split: train
        path: mixed_levels/*

DeepCaptcha Dataset: AI-Resistant CAPTCHA Benchmark

Overview

DeepCaptcha is a professional-grade dataset designed for benchmarking the AI-resistance and robustness of computer vision models. It contains CAPTCHA images generated with varying levels of adversarial protection, specifically engineered to thwart common automated recognition attacks (CNNs, OCR, etc.) while remaining human-readable.

Dataset Structure

The dataset is organized into 5 primary difficulty levels based on AI resistance techniques and PSNR (Peak Signal-to-Noise Ratio) degradation:

• Level 0 (Baseline): 600 images - No AI resistance. Standard CAPTCHAs.
• Level 1 (Basic): 600 images - Basic AI resistance (Target PSNR: 44.8dB).
• Level 2 (Moderate): 600 images - Moderate AI resistance (Target PSNR: 39.1dB).
• Level 3 (Advanced): 600 images - Advanced AI resistance (Target PSNR: 40.4dB).
• Mixed Levels: 600 images - A balanced mix of all resistance levels for general robustness training.

Total Images: 3,000 professional CAPTCHA samples.

Key Features

• Aesthetic Excellence: High-contrast, vibrant designs using premium typography.
• Adversarial Protection: Includes frequency-domain perturbations and progressive difficulty.
• Human-Centric: Maintains >95% human accuracy across all resistance levels.
• Machine Learning Ready: Includes a comprehensive metadata/ directory with JSON labels and an ML_USAGE_GUIDE.md.

Data Analysis

The dataset distribution covers:

• Text Length: 3 to 7 characters (Normal distribution centered at 5).
• Color Modes: Color (49.4%) and Black & White (50.6%).
• Character Frequency: Balanced distribution of alphanumeric characters.
• Image Dimensions: Optimized for standard input sizes (e.g., 250x120, 300x100).

Usage for ML Research

This dataset is ideal for:

1. Robustness Testing: Measuring how accuracy degrades as AI resistance increases.
2. Adversarial Training: Training models to be more resilient to frequency-domain noise.
3. OCR Benchmarking: Testing the limits of state-of-the-art OCR engines (Tesseract, PaddleOCR, etc.).

For detailed implementation patterns and evaluation metrics, please refer to the ML_USAGE_GUIDE.md included in the dataset.

License

This dataset is released under the MIT License.

Citation

If you use this dataset in your research, please cite the original project: DeepCaptcha GitHub Repository