uibk-uncover / knowyourlibrary

Systematic comparison of libjpeg versions 1998-2022

KnowYourLibrary

This is the code repository for the paper Know Your Library: How the libjpeg Version Influences Compression and Decompression Results [1]. It contains the codebase to replicate the systematic comparison of all libjpeg versions between 1998 and 2022.

Setup

Clone the repository to your computer

git clone https://github.com/uibk-uncover/KnowYourLibrary
cd KnowYourLibrary

Install the Python dependencies in a clean virtual environment. You might need to setup opencv.

python3 -m venv venv
source venv/bin/activate
pip install -r docker/requirements.txt

⚠️ Note that the installation takes a few minutes. This is because the libjpeg versions (part of package jpeglib) are source distributions to be compiled on your computer.

❗ You might need to setup opencv beforehand. To simplify, there is a Dockerfile that you can use. Follow the instructions below.

Docker

For convenience, the repo contains docker environment that you can use. Build the docker image by

docker build -t knowyourlibrary:latest -f docker/Dockerfile .

After 15 minutes, the image is built. Run the experiments with

docker run \
    --rm -it \
    --platform linux/amd64 \
    knowyourlibrary:latest \
    python src/run.py

The default data directories are /root/Datasets/*.

Usage

The test cases are executed with the following command.

python src/run.py

The script accepts the argument --help for usage instructions and availaible options.

python src/run.py
    [all|compression|decompression]
    [-i|--input "alaska=<path-to-alaska>;boss=<path-to-boss>"]
    [-n|--number <number-of-samples>]
    [--help|-h]

Using this interface, you can specify whether to test compression, decompression or both. You can also specify whether to run for colored (using the ALASKA dataset [2]) or grayscale images (using the BOSSBase dataset [3]). You can overwrite the default location with a custom one. In addition you can choose how many images from dataset will be used.

By default program uses at most 1000 images + certain specifically chosen (with maximal and minimal saturation, synthetic "checkerboard" with sharp edges etc.). The default location of the ALASKA dataset is ~/Datasets/ALASKA_v2_TIFF_256_COLOR, for BOSSBase it is ~/Datasets/BOSS_tiles.

Examples

To run compression test on ALASKA in directory /alaska and BOSSBase in default directory, type

python run.py compression -i "alaska=/alaska;boss"

To run decompression test on BOSSBase in directory /data/boss, but only use 30 images, type

python run.py decompression -i "boss=/data/boss" -n 30

There are example images in this repository. Execute tests on them only with

python run.py all -i "alaska=./data/alaska;boss=./data/boss" -n 15

Results

You can find examples of results for compression and decompression in results/ directory. Both contain baseline_color and baseline_grayscale directories with image files in naming format <filename>_<version>.[png|jpeg].

For example, data/alaska/00001.tif has been compressed with libjpeg 8 to results/compression/baseline_color/00001_8.jpeg. The same tif file was compressed with libjpeg 9e and decompressed with libjpeg 9b to results/decompression/baseline_color/00001_9b.png. All the images from results/ directory were created by results/create_baseline.py.

In the results/, you can find hashes of result examples in results/*.sha256. These were created by results/create_hash.sh.

Repository structure

The following files and directories contain the experiments.

• src/ = Python implementation
• src/run.py = entrypoint for executing
• data/*.sha256 = SHA256 hashes of the files
• log/ = example of logs, printed by src/run.py on stdout
• results/ = examples of baseline compression and decompression

The following files are intended to facilitate the repuducibility.

• requirements.txt = Python dependencies
• Dockerfile = Docker file
• data/alaska, data/boss = 15 example files from each dataset, to see limited results without need to get full datasets + 2 additional (the most and the least saturated)

References

[1] M. Benes, N. Hofer, and R. Böhme. 2022. Know Your Library: How the libjpeg Version Influences Compression and Decompression Results. In IH&MMSec. ACM, ?-?.

[2] R. Cogranne, Q. Giboulot, and P. Bas. 2019. The ALASKA steganalysis challenge: A first step towards steganalysis. In IH&MMSec. ACM, 125–137.

[3] P. Bas, T. Filler, and T. Pevný. 2011. Break our steganographic system. In IH (LNCS, Vol. 6958). Springer, 59–70.