DL Techniques for Generation of Steganographic Images with Hidden Audio
Welcome, Mr. Danyal, to the repository for the Deep Learning-Based Audio Steganography System. This project combines state-of-the-art deep learning techniques to securely hide audio within images and subsequently reveal it with exceptional fidelity. By integrating an advanced audio-to-image conversion process with a powerful deep learning model originally designed for image steganography, this system provides a robust, high-capacity solution for secure audio transmission, watermarking, and covert communications—all while preserving the visual quality of the cover image.
- Overview
- Process Flow
- Key Features
- System Architecture
- Deep Learning Model Reference
- Installation
- Usage
- Code Structure
- Repository Layout
- Contributing
- License
- Acknowledgements
This repository extends the concepts of deep image steganography to audio data. The system first converts an audio signal into a 256×256 RGB image, then leverages an advanced deep learning framework to embed this audio image into a cover image without compromising its visual quality. On the extraction side, the system recovers the hidden audio image from the stego-image and reconstructs the original audio waveform with minimal loss. This unique integration ensures that high-fidelity audio is securely concealed and later retrieved, making it an ideal tool for secure communications and digital watermarking.
-
Step 1: Audio-to-Image Conversion
- Normalization: Audio samples are normalized from a range of [-1, 1] to [0, 1].
- Reshaping: Data is padded or truncated to exactly 196,608 values (to form a 256×256×3 array).
- Scaling: Each normalized sample is scaled to an 8-bit intensity value (0–255), resulting in an image that encodes the audio waveform.
-
Step 2: Deep Learning-Based Audio Image Embedding
- A specialized deep learning model seamlessly embeds the generated audio image into a cover image using multi-layer feature fusion.
- The resulting stego-image is visually indistinguishable from the original cover image while securely containing the hidden audio data.
-
Step 3: Deep Learning-Based Audio Image Extraction
- The stego-image is processed through a trained reveal model that accurately recovers the embedded audio image.
-
Step 4: Image-to-Audio Conversion
- The recovered audio image is converted back into an audio waveform, preserving essential frequency and time-domain characteristics.
-
Step 5: Post-Processing and Audio Enhancement
- Advanced post-processing techniques (adaptive filtering, spectral rebalancing, and harmonic correction) are applied to remove noise and enhance audio clarity.
- A transmission degradation simulation is included to emulate real-world conditions and ensure robustness.
- High-Fidelity Audio Recovery: Retains complete audio information for lossless reconstruction.
- Imperceptible Embedding: The stego-image is visually identical to the cover image.
- Robust Deep Learning Models: Utilizes advanced neural networks for both embedding and extraction.
- Sophisticated Post-Processing: Enhances recovered audio by mitigating noise and distortions.
- User-Friendly Interface: A Streamlit-based GUI enables real-time interaction for seamless operation.
The system is divided into two primary components:
-
Image Embedding Phase:
- Converts the audio signal into an image representation.
- Embeds the audio image into a cover image using a deep learning model.
-
Image Extraction Phase:
- Extracts the hidden audio image from the stego-image.
- Reconstructs the original audio waveform with advanced post-processing enhancements.
This project incorporates the powerful deep learning model from the repository arnoweng/PyTorch-Deep-Image-Steganography. The original implementation is a PyTorch-based solution for image steganography, inspired by the paper "Hiding Images in Plain Sight: Deep Steganography". Key highlights of the original model include:
-
Hiding Network (H-net):
- An U-net structured convolutional network that takes a 6-channel tensor (concatenation of cover and secret images) and outputs a container image.
- Ensures that the container image is nearly identical to the cover image, with minimal perceptible differences.
-
Revealing Network (R-net):
- A convolutional network with six layers (each using a 3×3 kernel, followed by batch normalization and ReLU activations, except the last layer) designed to recover the hidden secret image from the container image.
- Achieves impressive recovery quality, with very low averaged pixel-wise discrepancies (APD) between the secret and the revealed secret.
Integration in Our Project:
In our adaptation, we extend this deep learning framework to handle audio data. The audio signal is first converted into an image, then processed by the H-net and R-net to embed and extract the hidden audio information. This integration not only preserves the high-fidelity nature of the original model but also introduces the capability to securely hide and reveal audio within images.
Visit the official PyTorch website to choose the version that matches your hardware.
Important: Always verify compatibility on the official site before installation.
Install the remaining libraries using pip:
pip install numpy librosa matplotlib soundfile torchvision pillow structlog streamlit| Library | Purpose |
|---|---|
| numpy | Numerical computations |
| librosa | Audio processing and transformation |
| matplotlib | Visualization and spectrogram display |
| soundfile | Audio file handling |
| torchvision | Image transformations using PyTorch |
| Pillow | Image processing and conversions |
| structlog | Structured logging for debugging |
| streamlit | GUI framework for real-time interaction |
- For CPU-only systems, install PyTorch with:
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu- Verify your installation with:
python3 -c "import torch; print('PyTorch Version:', torch.__version__)"Clone the repository and navigate to the project directory:
git clone https://github.com/danyalwg/DL-Audio-Stego.git
cd DL-Audio-StegoDownload the zipped checkpoint files from here and extract them into the checkPoint folder.
Launch the Streamlit GUI to begin the embedding and extraction process:
streamlit run streamlit_app.pyThis web-based interface provides an intuitive and interactive environment for processing audio steganography in real time.
| File Name | Functionality |
|---|---|
convert_to_image.py |
Converts an audio file into a 256×256 RGB image representation. |
embed_audio_image.py |
Embeds the audio image into a cover image using the deep learning model. |
extract_audio_image.py |
Extracts the hidden audio image from the stego-image. |
convert_to_audio.py |
Reconstructs the audio waveform from the recovered image. |
audio_processing.py |
Applies post-processing to enhance the recovered audio quality. |
streamlit_app.py |
Provides the GUI for user interaction with the system. |
DL-Audio-Stego/
│ audio_processing.py
│ convert_to_audio.py
│ convert_to_image.py
│ embed_audio_image.py
│ extract_audio_image.py
│ streamlit_app.py
│ README.md
│ requirements.txt
│
├───checkPoint
│ netH_epoch_73,sumloss=0.000447,Hloss=0.000258.pth
│ netR_epoch_73,sumloss=0.000447,Rloss=0.000252.pth
│
├───models
│ HidingUNet.py
│ RevealNet.py
│ __init__.py
│ __pycache__/
│
├───utils
│ transformed.py
│ __pycache__/
│
└───__pycache__/
We extend our sincere gratitude to the open-source community and especially to the maintainers of arnoweng/PyTorch-Deep-Image-Steganography for their pioneering work. Their innovative approach to deep learning-based steganography has been instrumental in the development of this audio steganography system.