yassineyousfi/alaska

Breaking the Alaska steganalysis challenge

🏔 ALASKA

This repo provides state-of-the-art pre-trained models for steganalysis in the JPEG domain, trained and used to win the ALASKA 1 steganalaysis challenge. Details about the architectures can be found in our paper.

Important Update

A new Alaska competition is now running on Kaggle, note that the settings are very different from the first edition of the competition: Image sizes, Quality factors, Embedding schemes, and Payload.

We have open-sourced our solution in this repo.

Features

• Color seperated feature maps extraction using pretrained SRNet models
• Arbitrary size steganalysis using pretrained detectors
• Notebooks to fine-tune feature extractors and train custom detectors
• Models are shared within the Tensorflow framework, and converted to ONNX for use with other deep learning frameworks.

Please note that shared models are only for JPEG quality factor 95.

Dependecies

Python 3.5+ and dependencies listed in requirements.txt. A Python3 compatible jpeg Package is included in the tools folder. System requirements: Mac OS, Linux (tested on Ubuntu 18.04)

Getting started - Downloading models

Please run the following python code to download the available models.

import requests
import zipfile
import os
home = os.path.expanduser("~")
user = home.split('/')[-1]

url = 'http://dde.binghamton.edu/download/alaska/models.zip'
local = home + '/alaska/models.zip'

r = requests.get(url)
with open(local, 'wb') as f:
    for chunk in tqdm(r.iter_content(chunk_size=2**10)): 
        if chunk:
            f.write(chunk)
with zipfile.ZipFile(local, 'r') as zipref:
    zipref.extractall(home + '/alaska/')
    
os.remove(local)

Getting started - Downloading datasets

This repo comes with minimal image examples, the complete datasets used to train these models have been removed from the official Alaska website by the organizers.

References

Please consider citing our paper if you find this repository useful.

@inproceedings{Yousfi2019Alaska,
 author = {Yousfi, Yassine and Butora, Jan and Fridrich, Jessica and Giboulot, Quentin},
 title = {Breaking ALASKA: Color Separation for Steganalysis in JPEG Domain},
 booktitle = {Proceedings of the ACM Workshop on Information Hiding and Multimedia Security},
 series = {IH\&\#38;MMSec'19},
 year = {2019},
 isbn = {978-1-4503-6821-6},
 location = {Paris, France},
 pages = {138--149},
 numpages = {12},
 url = {http://doi.acm.org/10.1145/3335203.3335727},
 doi = {10.1145/3335203.3335727},
 acmid = {3335727},
 publisher = {ACM},
 address = {New York, NY, USA},
} 

alaska's Issues

Python3 Jpeg Toolbox

I am having issues installing/compiling the JPEG Toolbox on my Windows 10 machine despite having all dependencies. Is there any way you could build wheel packages for the toolbox or if there is requirements that aren't listed I might not have installed.

MLP output

In this file: "tf_extract_features_color_separated.ipynb" which is for test, the variable stego_schemes includes only steganography methods, not "COVER":

stego_schemes = ['EBS', 'JUNI', 'NSF5', 'UED']

The variable "logits" as the output of test:

logits = sess.run(y, feed_dict={x: test_feature_map})

includes 5 numbers.

How "stego_schemes[np.argmax(softmax(logits))]" can give "COVER" as the output while we don't have it in the list in "stego_schemes "?

Now when I use Models (QF=95) and the image file "iPAD-pro-7.1-13inch_2065.jpg", I always get "JUNI" as result. It doesn't matter the input is COVER, EBS, JUNI, NSF5 or UED.

"ALASKA steganalaysis challenge" link in readme file throws error

http://alaska.utt.fr/alaska1 link throws 502 Bad Gateway Error

[Jpeglib] Code crash when call jpeglib.jpeg

After cloning the repo, I've compiled jpeglib on MacOS. I can import tools.python3_jpeg_toolbox as jpeglib. However when testing to call jpeglib.jpeg("temp.jpg"), I got the crash. Did you face it as well?

Error in training for 'Y'

The program 'tf_fine_tune_branch.ipynb' works fine for training 'YCrCb', but when I change it to 'Y', the program gives error:

InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shape= [3,3,3,64] rhs shape= [3,3,1,64]
[[node Assign_2 (defined at ..\tools\train_estimator.py:124) ]]