Official Implementation of InSeC – A deep learning framework for detecting Joint Parallel Steganography (JPS) in low-bit-rate VoIP compressed speech streams.
Published in IEEE Access, December 2024
InSeC (Inter‑codeword Sensitivity Caption) is a novel steganalysis model designed to detect hidden information in VoIP compressed speech, particularly when multiple steganographic algorithms (e.g., CNV and PSR) are used simultaneously – a scenario known as Joint Parallel Steganography (JPS). Unlike existing methods that target a single steganographic algorithm or suffer from poor generalization, InSeC introduces a two‑stage neural architecture that:
- Captures steganography‑sensitive codeword pairs from multiple perspectives (feed‑forward, stacked LSTMs, and channel‑wise attention).
- Re‑perceives fine‑grained correlations via a dedicated CNN module to amplify discriminative features.
Our method achieves state‑of‑the‑art detection accuracy across various embedding rates, speech durations, and packet loss conditions, while satisfying real‑time requirements (2.39 ms to process 1 second of speech).
- ✅ Multi‑algorithm steganalysis – Detects CNV, PSR, and their combination (JPS) in a single unified model.
- ✅ High accuracy – Outperforms four recent baselines by 9.07% ~ 25.27% on JPS detection (see results below).
- ✅ Robust to low embedding rates – Maintains >50% accuracy even at 1% embedding rate (substantially better than random).
- ✅ Short‑segment support – Works on speech frames as short as 0.1 seconds (excellent for real‑time streaming).
- ✅ Resilient to packet loss – Over 81% accuracy at 20% packet loss on JPS datasets.
- ✅ Lightweight & fast – Only 2.39 ms per 1‑second sample, suitable for backbone router deployment.
InSeC consists of three main modules:
-
Steganography‑Sensitive Codeword‑Pair Caption Module (SCM)
- Three parallel branches:
- Feed‑Forward Fully Connected (FFC) – models simple non‑linear mappings.
- Stacked LSTMs – captures both local and long‑range dependencies.
- Channel Interaction Attention Module (CIAM) – weights important feature channels.
- A fusion layer aggregates outputs from the three branches.
- Three parallel branches:
-
Fine‑Grained Correlation Re‑Perception Module (FCRM)
- Three 1D convolutional layers (kernel size 3) with batch normalization and ReLU.
- Average pooling + fully connected layer to produce a compact 32‑D feature vector.
-
Feature Classification Module (FCM)
- Two fully connected layers with dropout (p=0.5) and sigmoid activation.
- Outputs the probability of hidden information.
Figure 2 from the paper – the full network structure.
| Embedding Rate | RNN‑SM | SFFN | CSW | SANet | InSeC (Ours) |
|---|---|---|---|---|---|
| 10% | 52.39% | 68.21% | 67.50% | 67.14% | 71.42% |
| 30% | 61.30% | 75.00% | 77.50% | 77.29% | 86.57% |
| 50% | 74.52% | 85.27% | 85.70% | 84.26% | 92.54% |
InSeC consistently outperforms all baselines across low, medium, and high embedding rates.
| Method | Time (ms per 1‑second speech) |
|---|---|
| RNN‑SM | 0.77 |
| InSeC | 2.39 |
| SFFN | 2.49 |
| SANet | 3.27 |
| CSW | 3.99 |
✅ InSeC is among the fastest models while maintaining the highest accuracy.
| Method | Accuracy |
|---|---|
| RNN‑SM | 52.07% |
| SFFN | 68.55% |
| CSW | 69.42% |
| SANet | 70.31% |
| InSeC | 86.99% |
InSeC excels even on extremely short utterances, crucial for real‑time streaming interception.
| Packet Loss | RNN‑SM | SFFN | CSW | SANet | InSeC |
|---|---|---|---|---|---|
| 5% | 73.22% | 79.40% | 81.21% | 82.75% | 88.25% |
| 15% | 71.01% | 77.34% | 80.11% | 81.23% | 84.74% |
- Python 3.8+
- PyTorch 1.9+
- CUDA (optional, for GPU acceleration)
git clone https://github.com/zhousandeqingshu/ZhCode.git
cd ZhCode
pip install -r requirements.txtpython train.py --stego_type JPS --embed_rate 30 --frame_len 10 --epochs 50python test.py --model_path checkpoints/best_model.pth --data_path ./test_dataThe model expects G.729 encoded speech in PCM format.
Use the dataset from Lin et al. (2018) containing:
- 41 hours of Chinese speech
- 72 hours of English speech
Generate cover/stego samples with CNV, PSR, or JPS embedding at desired rates (10%–100%).
Frame length can be set from 0.1 s to 2 s (default: 10 s for training).
python train.py --stego_type JPS --embed_rate 30 --frame_len 10 --epochs 50
python test.py --model_path checkpoints/best_model.pth --data_path ./test_data
If you find our work useful for your research, please cite:
@article{Zhang2024InSeC,
author = {Hao Zhang and Jie Yang and Feipeng Gao and Jiacheng Yuan},
title = {InSeC: Steganalysis Model Based on Inter-Codeword Sensitivity Caption for Compressed Speech Streams},
journal = {IEEE Access},
volume = {12},
pages = {192252--192262},
year = {2024},
doi = {10.1109/ACCESS.2024.3519094},
note = {Accepted 9 December 2024, published 17 December 2024}
}
This project is licensed under the MIT License – see the LICENSE file for details.
Issues and pull requests are welcome! For major changes, please open an issue first to discuss what you would like to change.
- First author: Zhang Hao – 3181280664@qq.com
- Code repository: https://github.com/zhousandeqingshu/ZhCode
If InSeC has helped your research, a ⭐ on GitHub would be an enormous encouragement to us and helps more people find this work！ https://api.star-history.com/svg?repos=zhousandeqingshu/ZhCode&type=Date