Effect of libjpeg version on CSM

This is the code repository for the paper The Effect of the JPEG Implementation on the Cover-Source Mismatch Error in Image Steganalysis [1]. It contains the codebase to replicate the CSM caused by a mismatch in libjpeg versions, which was described in our previous paper [2].

Setup

This codebase uses Python and Matlab.

Python

The code was executed on Python 3.8.11. Install dependencies with

pip3 install jpeglib==0.10.11 pillow==9.0.0 numpy==1.19.5

Matlab

Matlab is executed locally in a Matlab environment.

You need the Matlab JPEG Toolbox by Phil Sallee. Prebuilt versions have been published, or you can build it from source. In that case, you need to include and link libjpeg.

mex -I/usr/local/include -I./include/jpegtbx include/jpegtbx/jpeg_read.c -L/usr/local/lib -ljpeg -v
mex -I/usr/local/include -I./include/jpegtbx include/jpegtbx/jpeg_write.c -L/usr/local/lib -ljpeg -v

In Matlab, you need to add alice/, eve/ and include directories to path.

Usage

We use the Python package jpeglib that allows the selection of the libjpeg version for compression and decompression. However, for J-UNIWARD embedding, JSRM feature extraction and EFLD model training and evaluation we use the implementation of DDE/Jessica Fridrich [3].

Alice: Downloading data

As pre-compressed covers $x_{\text{src}}$ we use the ALASKA2 dataset [4]. As it has around 16GB, you will have to download it directly from the original repository of the ALASKA2 steganalysis challenge.

Downloading can be done by

bash data/ALASKA_v2_TIFF_256_COLOR.sh

The script data/ALASKA_v2_TIFF_256_COLOR.sh was provided by the authors of the ALASKA2 steganalysis challenge. The dataset is stored in data/ALASKA_v2_TIFF_256_COLOR.

Alice: Cover preparation

Once you downloaded the data, you have to compress and decompress it by different versions for Alice. In this step, 10,000 random data samples are drawn.

# perform Alice using 6b
python alice/create_covers.py 6b
# perform Alice using 7
python alice/create_covers.py 7

The compressed and decompressed images are stored in data/ALASKA_6b for version 6b and data/ALASKA_7 for version 7.

Alice: Embedding

Execute the Matlab code from alice/create_stego.m to create stego images from the covers with J-UNIWARD at rate 0.4bpnzac. The outputs are stored in data/ALASKA_*/stego_juniward_0.4.

Eve: Image preparation

As an input, Eve uses data/ALASKA_*/cover and data/ALASKA_*/stego_juniward_0.4, so Alice' part needs to be performed before this step.

Compress and decompress the input with different libjpeg versions for Eve.

# perform Eve using 6b and 7, when Alice used 6b
python eve/create_covers.py 6b
python eve/create_stego.py 6b
# perform Eve using 6b and 7, when Alice used 7
python eve/create_covers.py 7
python eve/create_stego.py 7

The compressed and decompressed images are stored in data/ALASKA_*/cover_compressed6b and data/ALASKA_*/cover_decompressed6b for Eve using version 6b and data/ALASKA_*/cover_compressed7 and data/ALASKA_*/cover_decompressed7 for Eve using version 7.

Eve: Feature extraction

Execute the Matlab code from eve/extract_jsrm.m to extract JSRM features from the input files.

This will output the following MAT files in data/ for cover and stego objects.

• data/C_jsrm_6b_6b.mat
• data/C_jsrm_7_7.mat
• data/C_jsrm_6b_7.mat
• data/C_jsrm_7_6b.mat
• data/S_jsrm_6b_6b_0.4.mat
• data/S_jsrm_7_7_0.4.mat
• data/S_jsrm_6b_7_0.4.mat
• data/S_jsrm_7_6b_0.4.mat

These MAT files contain either C_jsrm or S_jsrm objects.

Eve: Model training

Execute the Matlab code from eve/train_model.m to train the model. Trained models are saved in the directory model/.

• model/model_6b_6b_0.4.mat
• model/model_7_7_0.4.mat
• model/model_6b_7_0.4.mat
• model/model_7_6b_0.4.mat

Eve: Model evaluation

Execute the Matlab code from eve/evaluate_model.m to evaluate the model and print its performance.

Repository structure

The repository has the following structure.

• alice = codebase for Alice
• eve = codebase for Eve
• include = shared codebase
• data = script for downloading data, destination for generated images
• model = trained models

Reference

[1] Beneš, M., Hofer, N., and Böhme, R. The Effect of the JPEG Implementation on the Cover-Source Mismatch Error in Image Steganalysis. In Proceedings of the 30th European Signal Processing Conference (EUSIPCO). EURASIP, Belgrade, Serbia, 2022. [PDF] [Video]

[2] Beneš, M., Hofer, N., and Böhme, R. Know Your Library: How the libjpeg Version Influences Compression and Decompression Results. In Proceedings of the 10th ACM Workshop on Information Hiding and Multimedia Security (IH&MMSEC). ACM, 2022, pp. 19–25. [PDF] [Publisher]

[3] Digital Data Embedding Laboratory. Department of Electrical and Computer Engineering. Binghamton University. [URL]

[4] Cogranne, R., Giboulot, Q., and Bas, P. The ALASKA Steganalysis Challenge: A First Step Towards Steganalysis “Into The Wild. In Proceedings of the ACM Workshop on Information Hiding and Multimedia Security (IH&MMSec). ACM, 2019, pp. 125-137. [PDF] [Publisher]