📸 Source Camera Model Identification via Federated Learning using Laplacian-based Patches

---

📖 Overview

Official implementation of the paper:

"Source Camera Model Identification via Federated Learning using Laplacian-based Patches"
Published in IEEE Transactions on Artificial Intelligence (IEEE TAI).

This work proposes a Federated Learning (FL) based framework for Source Camera Model Identification (SCMI) using Laplacian-based image patches and the proposed FedFFTNet architecture.

The framework consists of:

• 📌 Laplacian-based informative patch extraction
• 🌐 Federated data distribution across clients
• 🧠 Federated model training using FedFFTNet
• 📷 Source Camera Model Identification

---

🔗 Paper

📄 Paper Link:
Source Camera Model Identification via Federated Learning using Laplacian-based Patches

---

📂 Repository Workflow

The complete pipeline consists of two major stages:

Raw Images
    │
    ▼
Laplacian-based Patch Extraction
    │
    ▼
Federated Client Distribution
    │
    ▼
FedFFTNet Training
    │
    ▼
Source Camera Model Prediction

---

📁 Dataset Structure

Ensure your dataset is pre-split into train and test folders.

data/
├── train/
│   ├── class_1/
│   ├── class_2/
│   └── ...
│
├── test/
│   ├── class_1/
│   ├── class_2/
│   └── ...

---

🧩 Step 1 — Laplacian-based Image Patching

Notebook:

Fed_DataSplit_Laplacian.ipynb

This notebook performs:

• Laplacian-based informative patch extraction
• Federated client data distribution
• Global server data allocation
• Test set generation

---

⚙️ Hyperparameters

Variable	Type	Default	Description
datapath	str	./data	Source dataset directory
number_clients	int	5	Number of federated clients
initial_ratio	float	0.06	Ratio of globally available data
dist_type	str	equal	Data distribution type (equal, random, skew)

---

🖼️ Patching Pipeline

Figure 1. Laplacian-based patch extraction and federated data distribution.

---

📦 Output Directory Structure

After execution, the generated patches/ directory should look like:

patches/
├── initial/
├── clients/
│   ├── client_1/
│   ├── client_2/
│   └── ...
│
└── test/

---

🧠 Step 2 — Federated Model Training

Notebook:

FedLearn_FedFFTNet.ipynb

This notebook performs:

• Federated training
• Global aggregation
• Local client optimization
• Source Camera Model classification

---

⚙️ Training Configuration

Variable	Type	Default	Description
root	str	./patches	Patched dataset directory
global_available	bool	True	Availability of global server data
batchsize	int	128	Training batch size
ROUNDS	int	100	Number of communication rounds
CLIENT_EPOCHS	int	1	Local client epochs per round
GLOBAL_EPOCHS	int	30	Global server training epochs
learning_rate	float	0.001	Optimizer learning rate
device	str	cuda:0	Training device

---

🏗️ FedFFTNet Architecture

Figure 2. Proposed FedFFTNet architecture for federated SCMI.

---

🚀 Running the Pipeline

Step 1 — Generate Patches

Run:

Fed_DataSplit_Laplacian.ipynb

This creates the federated patch dataset.

---

Step 2 — Train FedFFTNet

Run:

FedLearn_FedFFTNet.ipynb

This starts federated training across all clients.

---

🧪 Supported Data Distribution Modes

Distribution	Description
equal	Equal data distribution across all clients
random	Random client-wise data distribution
skew	Non-IID skewed distribution

---

📚 Citation

If you use this work in your research, please cite:

@article{chakraborty2025source,
  title={Source Camera Model Identification via Federated Learning using Laplacian-based Patches},
  author={Chakraborty, Rejoy and Goyal, Puneet},
  journal={IEEE Transactions on Artificial Intelligence},
  year={2025},
  publisher={IEEE}
}

---

⚠️ Disclaimer

This repository is intended strictly for:

• Academic research
• Educational purposes
• Experimental evaluation

---

👨‍💻 Authors

• Rejoy Chakraborty
• Puneet Goyal

---

⭐ Acknowledgement

If you find this repository useful, consider giving it a ⭐ on GitHub.

---