---
license: mit
size: 100k
task_categories:
- image-classification
tags:
- steganography
pretty_name: BOSS-Based Cropped Steganography Dataset
size_categories:
- 100K<n<1M
---

# Dataset Card: BOSS-Based Cropped Steganography Dataset

## Dataset Overview
- **Name**: *BOSS-Based Cropped Steganography Dataset* (working title)
- **Description**:
  This dataset is a prepared and cropped version of the BOSSbase v1.01 dataset, commonly used in steganography and steganalysis research. Each original 512×512 grayscale image is split into four non-overlapping 256×256 patches to support deep learning experiments with smaller, manageable image sizes. The dataset supports binary classification tasks such as cover vs. stego detection.
- **Purpose**:
  Designed to enable reproducibility of research in image steganalysis.
- **Supported Tasks**:
  - Binary classification (cover vs. stego)
  - Steganalysis model evaluation
- **Author**: *Italo Amaya*
- **License**: MIT License
- **Intended Use**: Academic research only

## Dataset Structure
- **Total Images**:
  - 20,000 **cover** images
  - 20,000 **WOW stego** images
  - 20,000 **S-UNIWARD stego** images
  *(Each stego set is embedded from the same 20,000 cover images, using different algorithms and randomized keys.)*
- **Image Format**: `.png`
- **Image Size**: 256×256 pixels
- **Color**: Grayscale (1 channel)
- **Directory Layout**:
  ```
  cropdataset/
  ├── cover/
  └── stego/
      ├── S-UNIWARD/
      │   └── 0.4bpp/
      │       └── stego/
      └── WOW/
          └── 0.4bpp/
              └── stego/
  ```
- **Labels**: Not explicitly stored — labels should be inferred during data loading (`cover = 0`, `stego = 1`)

## Data Generation Process
- **Original Dataset**: [BOSSbase v1.01](https://dde.binghamton.edu/download/ImageDB/BOSSbase_1.01.zip)
- **Cropping Script (Python)**:
  ```python
  from PIL import Image
  import os
  from tqdm import tqdm

  def crop_images(input_folder, output_folder, patch_size=(256, 256)):
      os.makedirs(output_folder, exist_ok=True)
      for filename in tqdm(os.listdir(input_folder), desc="Cropping"):
          try:
              input_path = os.path.join(input_folder, filename)
              if os.path.isfile(input_path):
                  img = Image.open(input_path)
                  if img.size == (512, 512):
                      w, h = patch_size
                      patches = {
                          "top_left": img.crop((0, 0, w, h)),
                          "top_right": img.crop((512-w, 0, 512, h)),
                          "bottom_left": img.crop((0, 512-h, w, 512)),
                          "bottom_right": img.crop((512-w, 512-h, 512, 512)),
                      }
                      for key, patch in patches.items():
                          patch.save(os.path.join(output_folder, f"{os.path.splitext(filename)[0]}_{key}.png"))
          except Exception as e:
              print(f"Error processing {filename}: {e}")
  ```
- **Embedding Tools**:
  > [!NOTE]
  > The embedding was done using random keys and as such this dataset is a random key dataset
  - [WOW](https://dde.binghamton.edu/download/stego_algorithms/)
  - [S-UNIWARD](https://dde.binghamton.edu/download/stego_algorithms/)
  *Both are official C++ implementations from the University of Binghamton.*
- **Embedding Parameters**:
  - Randomized keys per image
  - 0.4 bits per pixel (bpp)

## Preprocessing & Transformations
- **Cropping**: Non-overlapping 256×256 patches from 512×512 originals
- **Augmentation**: None
- **Normalization**: None — images retain original pixel value ranges

## Citation and Attribution
- **GitHub Repository**: *[Link to be added once public]*
- **Associated Paper**: *[Placeholder for paper title or DOI]*
  *Please cite this work when using the dataset.*