🚨 Advanced Geometric Aggregation and Gradient Regularization Framework for Semi-supervised Federated Intrusion Detection in Heterogeneous IoT Environments

This repository contains the implementation of an Enhanced Semi-Supervised Federated Learning-based Intrusion Detection System (USSFL-IDS). It is based on the paper:

📄 Zhao et al., "Semi-supervised Federated-Learning-Based Intrusion Detection Method for Internet of Things", IEEE 2023

We further extend their work by integrating robust aggregation, gradient variance tracking, Fishr regularization, and visual analytics to improve generalization and transparency.

---

📌 Project Highlights

🔐 Problem

Detecting network intrusions in Internet of Things (IoT) environments using semi-supervised federated learning, where client data is heterogeneous (non-IID) and privacy-sensitive.

🚀 Key Enhancements

Feature	Description
Gradient Aggregation	Supports both Arithmetic and Geometric Mean
Fishr Regularization	Penalizes gradient variance divergence across clients
Gradient Variance	Tracks gradient dispersion for robust training
Discriminator Training	Differentiates known vs unknown samples
Voting Mechanism	Aggregates logits using hard label voting
Visualizations	Confusion Matrix, Class Distribution, Accuracy/F1 Trends

---

📂 Directory Structure

USSFL/
├── utils/
│   ├── helper.py              # Gradient aggregation & Fishr loss
│   ├── model_utils.py         # CNN model architecture
│   ├── creator.py             # Dataset creation and client/server setup
│   ├── train_utils.py         # Training, prediction, evaluation utilities
│   ├── process_data_utils.py  # Data preprocessing, splitting
│   └── visualization.py       # 📊 All plots for metrics & analysis
├── data/
│   └── nba_iot_1000/          # N-BaIoT preprocessed CSVs per device & attack type
├── USSFL-IDS.py               # Main training script
└── README.md

---

🧠 Core Techniques

• Federated Learning (FL) across IoT devices
• Semi-supervised Learning (SSL) with open and private sets
• Dirichlet distribution-based split for simulating realistic non-IID scenarios
• Fishr Regularization: Aligns gradient variance to minimize distributional shift
• Hard-label voting aggregation from client logits

---

📈 Visualizations

Plot Type	Generated By
Client-Class Distribution	plot_class_distribution()
Accuracy, F1, Precision Curves	plot_metrics()
Confusion Matrix	plot_conf_matrix()
Comm. Overhead vs Accuracy	plot_comm_overhead_vs_accuracy()
θc Threshold Impact	plot_theta_accuracy()
Label Strategy Comparison	plot_label_strategy_comparison()

All graphs are generated after the final communication round.

---

🛠 Setup Instructions

✅ Prerequisites

• Python 3.8+
• PyTorch
• NumPy
• Scikit-learn
• Matplotlib

📦 Install Dependencies

pip install -r requirements.txt

📁 Dataset

Place preprocessed N-BaIoT CSVs under:

data/nba_iot_1000/<DeviceName>/<AttackType>_train.csv
data/nba_iot_1000/<DeviceName>/<AttackType>_test.csv

---

▶️ Running the Project

python USSFL-IDS.py

This will:

1. Load the dataset.
2. Create clients and server.
3. Train for configured rounds.
4. Display visual metrics.

---

📊 Example Output

• Final test accuracy
• F1-score and precision plots
• Confusion matrix for test predictions
• Communication overhead calculations
• Class-wise client distribution

---

🧪 Evaluation Metrics

• Accuracy
• F1-Score (Macro)
• Precision (Macro)
• Confusion Matrix
• Communication Cost

---

📚 Reference

• Zhao et al., "Semi-supervised Federated-Learning-Based Intrusion Detection Method for Internet of Things", IEEE 2023
• This implementation includes original enhancements like Fishr loss and improved aggregation.

---

✍️ Author

• Your Name — LinkedIn | GitHub

---

📌 License

This project is for academic and research use only.

---