B.Tech Project — Extends the SecureDSC framework by Shi et al. (IEEE Communications Letters, March 2025) with physical-layer key generation and adaptive adversarial training.
This project implements and enhances a secure wireless semantic communication system where:
- Alice (transmitter) sends text messages to Bob (receiver) over a noisy wireless channel
- Eve (eavesdropper) intercepts the same channel symbols but cannot reconstruct the content
- Security is achieved through adversarial training with an integrated encryption module
| # | Enhancement | Replaces | Impact |
|---|---|---|---|
| 1 | CSI-Based Dynamic Key Generation | Random session keys | Eliminates key distribution bottleneck |
| 2 | Adaptive λ Scheduler | Fixed λ = 6 | Self-tuning adversarial loss, +5.2% Bob BLEU-1 |
graph LR
CSI["CSI Channel Estimates<br/>(h_A ≈ h_B by reciprocity)"] --> KG["CSI Key Generator<br/>★ Enhancement 1"]
KG --> KEY["key_A / key_B"]
A["m (Source Text)"] --> B["Semantic Encoder<br/>4× Transformer Layers"]
B --> C["Encryptor<br/>(key_A from CSI)"]
C --> D["Channel Encoder"]
D --> E["AWGN Channel"]
E --> F["Channel Decoder (Bob)"]
F --> G["Decryptor<br/>(key_B from CSI)"]
G --> H["Semantic Decoder<br/>4× Transformer Layers"]
H --> I["m̂ (Reconstructed)"]
E --> J["Channel Decoder (Eve)"]
J --> K["Decryptor<br/>(random key — no CSI)"]
K --> L["Semantic Decoder (Eve)"]
L --> M["m̄ (Eve's attempt)"]
SCHED["Adaptive λ Scheduler<br/>★ Enhancement 2"] -.-> LOSS["Joint Loss Computation"]
For a comprehensive breakdown of the training process, evaluation metrics (including BLEU scores and Key Agreement Rate), and a detailed comparison with the base paper, please refer to the Project Analysis & Comparison Document.
Enhanced-SecureDSC/
├── model.py # Full model architecture (472 lines)
│ ├── SemanticEncoder / Decoder # 4× Transformer layers, d_model=128
│ ├── Encryptor / Decryptor # Key-conditioned encryption modules
│ ├── ChannelEncoder / Decoder # Dense JSCC layers
│ ├── CSIKeyGenerator # ★ MLP: ℂ^64 → {-1,+1}^64
│ ├── AdaptiveLambdaScheduler # ★ Bang-bang controller
│ └── SecureDSC # End-to-end system class
├── train.py # 4-phase adversarial training loop
├── evaluate.py # BLEU scoring + key agreement evaluation
├── preTokenize.py # EuroParl pre-tokenization script
├── requirements.txt # Dependencies (PyTorch, Transformers)
├── eval_results.json # Final evaluation metrics
├── training_history.json # Epoch-by-epoch training logs
└── project_analysis_and_comparison.md # Detailed technical analysis
- Python 3.8+
- NVIDIA GPU with CUDA 12.1+ (recommended)
# Clone the repository
git clone https://github.com/krish-kedia/Enhanced-SecureDSC.git
cd Enhanced-SecureDSC
# Create and activate virtual environment
python -m venv venv
# Windows:
.\venv\Scripts\activate
# Linux/Mac:
source venv/bin/activate
# Install dependencies
pip install -r requirements.txt# Full training (150 epochs, ~1.96M sentences)
python train.py --epochs 150 --snr 12 --batch_size 512
# Quick test run
python train.py --epochs 5 --snr 12 --batch_size 64 --dataset_size 500python evaluate.py --model_path securedsc_enhanced.pt --snr_range 0 3 6 9 12 15- 4-phase alternating schedule: Semantic codec → Encryption/Decryption → Full joint (adversarial) → Eve independently
- Optimizer: Adam (lr=2.5×10⁻⁴, β₁=0.9, β₂=0.98)
- Dataset: EuroParl EN-FR (English side), seq_len=20, 90/10 train/test split
- Tokenizer: BERT-base-uncased (30,522 vocab)
- Gradient clipping: max_norm=1.0
If you use this work, please cite the base paper:
@article{shi2025secure,
title = {Secure Transmission in Wireless Semantic Communications With Adversarial Training},
author = {Shi, Jiting and Zhang, Qianyun and Zeng, Weihao and Qin, Zhijin},
journal = {IEEE Communications Letters},
volume = {29},
number = {3},
pages = {487--491},
year = {2025},
publisher = {IEEE}
}