Audio Steganalysis with Deep Learning
Overview
The tf_audio_steganalysis project is focused on analyzing audio steganography using deep learning techniques, specifically with TensorFlow. Led by Yuntao Wang, this project provides a platform for users to design their own neural networks to detect and analyze hidden data within audio files, particularly MP3s. The work has earned recognition, including the Best Paper Award at IH&MMSec 2018.
Key Papers and Contributions
The project highlights several important papers:
-
CNN-based Steganalysis of MP3 Steganography in the Entropy Code Domain - This work received the Best Paper Award at IH&MMSec 2018. It explores using Convolutional Neural Networks (CNNs) to detect hidden information in MP3 files using entropy code domain analysis. The paper is accessible via ACM and a pdf version is available.
-
RHFCN: Fully CNN-based Steganalysis of MP3 with Rich High-Pass Filtering - Presented at ICASSP 2019, this research extends the use of CNNs with high-pass filtering to improve detection accuracy. You can read more through the IEEE link and the pdf.
-
MP3 Steganalysis based on Joint Point-wise and Block-wise Correlations - This study, published in Information Sciences, focuses on using joint correlations to enhance steganalysis. It’s detailed in the Elsevier journal.
Required Packages
To make full use of the project, several packages are necessary, including TensorFlow, numpy, pandas, matplotlib, and others. Users should execute pip install -r requirements.txt to install these. For new environments, creating a virtual environment is recommended to avoid conflicts.
Instructions for Use
-
Installation: Start by installing Python 3.x or Anaconda and ensure it’s added to your system's environment variables.
-
Configure GPU: For those with GPU capability, configuring the runtime is optional but can accelerate the training process.
-
Install Dependencies: Use the requirements.txt file provided to install all dependencies promptly.
-
Run the Code: Follow the example configuration files to execute the code.
-
Visualization: Utilize TensorBoard to visualize data during training. This includes tracking the accuracy and loss metrics with the command
tensorboard --logdir=/path of log. -
Custom Network Design: If developing a custom network, refer to the provided instruction guide.
-
Development Environment: All source code is compatible with Pycharm, with recommendations to set the hard wrap at 180 to avoid editor warnings.
Project Files
The project comprises several directories and files:
- audio_samples: Contains audio samples for testing and analysis.
- data_processing: Tools for dataset creation and QMDCT coefficient extraction.
- jupyter: A folder for debugging using Jupyter notebooks.
- papers: Includes papers, presentations, datasets, and introductions to recent works.
- setup: Provides setup and configuration files.
- src: Contains the project's source code.
Conclusion
This project serves as a comprehensive resource for exploring audio steganalysis with advanced machine learning techniques. It offers extensive tools and documentation for researchers and developers looking to delve into the field of digital data hiding and detection in audio files.