Database

LSSD

Large Scale Steganalysis Database — 2 million 256×256 JPEG images (colour or grayscale) built for deep-learning steganalysis, “into the wild”, under controlled conditions.

Overview of the LSSD database: mosaic of 256×256 images
fig.00 — Overview of the LSSD database: 256×256 images, colour and grayscale.

Access

Three ways to get the database and its various sizes:

Script
download_lssd

Download script LSSD_download_script.sh — select by database, size, format and colour.

 Dedicated site
lssd.lirmm.fr

Portal dedicated to the database, hosted by LIRMM.

 GitHub
Yiouki / download_lssd

Repository with the script, folder structure and RAW-development tools.

Database content

LSSD images are developed from RAW images (sensor, before processing), then cropped and re-compressed to 256×256 JPEG.

Diagram of the process developing RAW images into 256×256 JPEG
fig.01 — Developing RAW images into 256×256 JPEG crops.
Image 3786 developed, ALASKA database (grayscale)
ALASKA · 3786
grayscale
Image 6456 developed, BOSS database (grayscale)
BOSS · 6456
grayscale
Image 51336 developed, ALASKA database (colour)
ALASKA · 51336
colour
Image developed, Wesaturate database (colour)
Wesaturate
colour

These images come from six public RAW image databases:

RAW source databaseImagesShare
ALASKA 280,00562.88%
Stego App24,12018.96%
BOSS10,0007.66%
RAISE8,1566.41%
Wesaturate3,6482.87%
Dresden1,4911.23%

Developed in cover and stego versions (JUNIWARD, 0.2 bpnzac), in JPEG and MAT formats, colour or grayscale, and in several sizes:

10K 50K 100K 500K 1M 2M TEST

Example download command:

sh LSSD_download_script.sh -b LSSD_2M -t JPEG -c Color -n Cover

Purpose

LSSD was designed to study steganalysis with very large training databases and under controlled conditions. Controlling the development of RAW images (demosaicing, resizing, JPEG re-compression) makes it possible to isolate the factors that truly affect neural-network performance, and to get closer to realistic “into the wild” conditions.

Citation

If you use LSSD, please cite:

H. Ruiz, M. Yedroudj, M. Chaumont, F. Comby, G. Subsol, “LSSD: a Controlled Large JPEG Image Database for Deep-Learning-based Steganalysis “into the Wild””, ICPR 2021 Workshop MMForWILD, LNCS 12666, Springer, January 2021. 
@inproceedings{ruiz2021lssd,
  title     = {LSSD: a Controlled Large JPEG Image Database for
               Deep-Learning-based Steganalysis "into the Wild"},
  author    = {Ruiz, Hugo and Yedroudj, Mehdi and Chaumont, Marc
               and Comby, Fr\'ed\'eric and Subsol, G\'erard},
  booktitle = {Pattern Recognition. ICPR International Workshops
               and Challenges (MMForWILD 2021)},
  series    = {LNCS},
  volume    = {12666},
  publisher = {Springer},
  year      = {2021},
  doi       = {10.1007/978-3-030-68780-9_38}
}

Resources

DOI ↗ Paper PDF ↓ Slides ↓ Video ↗ All publications →