📡 Jammer Detection & Classification in Wireless Communications

CNN-based detection and classification of jamming signals in 5G wireless systems using synthetically generated heatmap datasets.

---

📌 Project Overview

This project develops and evaluates Convolutional Neural Networks (CNNs) to detect and classify jamming signals in wireless communication systems. The model can:

• Detect whether a communication signal is being jammed (binary classification)
• Identify the specific type of jammer present (multi-class classification)

The full pipeline covers synthetic dataset generation, data augmentation, CNN model design, training, evaluation, and improvement analysis.

---

🔬 Pipeline at a Glance

Sionna Simulation
      ↓
Generate Heatmap Images (133×14)
      ↓
5G Scheduler Masking (Data Augmentation)
      ↓
Build Binary & Multi-Class Datasets
      ↓
Train PaperCNN Model
      ↓
Evaluate & Analyze
      ↓
Improvement Recommendations

---

📂 Project Stages

1. 🗂️ Dataset Generation

Synthetic heatmap images were generated using Sionna, NVIDIA's open-source TensorFlow-based library for link-level simulations.

Resource Grid Parameters:

Parameter	Value
Physical Resource Blocks (PRBs)	133
Subcarriers per PRB	12
FFT Size	1596
OFDM Symbols	14
Subcarrier Spacing	30 kHz
Modulation	16-QAM
Channel Model	TDL-C
Delay Spread	300 ns
Carrier Frequency	3.5 GHz
Fixed SNR	10 dB

Jammer Types (5 Classes):

Class	Description
🟢 Noise	Standard AWGN — no jamming signal
🔴 Barrage	Wideband, continuous jamming across a large spectrum portion
🟠 Sweep	Jamming signal that sweeps across frequencies over time
🟡 Spot	Narrowband, continuous jamming fixed at one frequency
🔵 Pulse	Intermittent jamming appearing for short bursts

Generation Process (per sample):

1. Define jammer-specific parameters (bandwidth, offset, width, spacing, etc.)
2. Fill a ResourceGrid with jammer symbols
3. Scale signal power using a random JNR value (−10 dB to +10 dB)
4. Apply TDL channel model + AWGN noise
5. Convert received signal → 133×14 heatmap (power averaged per PRB & OFDM symbol)

Dataset Sizes:

Dataset	Description	Samples
multi_dataset	All 5 classes (Noise, Barrage, Sweep, Spot, Pulse)	10,500
binary_dataset	Not-jammed (Noise) vs. Jammed (all 4 jammers)	Balanced

5 classes × 21 JNR values × 100 samples = 10,500 initial samples

---

2. 🔄 Data Augmentation

To simulate real-world conditions where received signals can be incomplete or partially blocked, a 5G scheduler mask was applied to a subset of heatmaps.

• Randomly selects contiguous blocks of PRBs and OFDM symbols
• Sets selected values to zero, mimicking unallocated or obstructed resource blocks
• Increases model robustness to data incompleteness

---

3. 🧠 Model Architecture — PaperCNN

A custom CNN architecture (PaperCNN) was designed for both classification tasks:

• Input: 133×14 heatmap images
• Task A: Binary classification → jammed / not jammed
• Task B: Multi-class classification → Noise / Barrage / Sweep / Spot / Pulse

---

4. 📊 Training & Evaluation

Models were trained on the generated and augmented datasets. Training trends were analyzed for:

• Accuracy and loss curves
• Overfitting detection
• Per-class performance

---

5. 🔍 Analysis & Improvement Suggestions

Post-training analysis identified potential issues including overfitting and proposed concrete improvement steps such as:

• Regularization techniques (Dropout, L2)
• Learning rate scheduling
• Expanding dataset diversity
• Cross-validation strategies

---

⚙️ Getting Started

Run in Google Colab (Recommended)

The notebook is large and may not render directly on GitHub. Click the badge below to open it directly in Colab:

Dependencies

pip install sionna tensorflow numpy matplotlib

Note: Sionna requires TensorFlow. A GPU runtime is strongly recommended in Colab (Runtime → Change runtime type → T4 GPU).

---

📦 Key Libraries

Library	Purpose
sionna	Link-level wireless simulation & heatmap generation
tensorflow / keras	CNN model building and training
numpy	Numerical operations and data handling
matplotlib	Heatmap and training curve visualization

---

📈 Results Summary

Task	Model	Notes
Binary Detection	PaperCNN	Jammed vs. Not Jammed
Multi-Class Classification	PaperCNN	5 jammer types

Detailed results, confusion matrices, and training curves are available inside the notebook.

---

🧪 Dataset Summary

JNR Range:     -10 dB to +10 dB (21 values)
Samples/class: 100 per JNR value
Heatmap size:  133 × 14 (PRBs × OFDM symbols)
Total samples: 10,500 (multi) | Balanced (binary)
Augmentation:  5G Scheduler Masking

---

📄 License

This project is licensed under the MIT License.

---

Built with Sionna, TensorFlow, and Python. Designed for research in 5G physical-layer security.