D4-Net: Detecting Deepfakes using a Dual-branch Deep Learner

📌 Overview

The D4-Net project introduces a dual-branch deep neural network for robust deepfake detection. It leverages two complementary pathways:

• Semantic Branch (RGB/Xception) → captures high-level spatial and facial cues.
• Frequency Branch (MesoNet/FFT) → extracts subtle frequency irregularities.

The outputs are integrated via cross-attention and adaptive fusion, leading to highly discriminative features for classification.

We evaluate the model on FaceForensics++ (FF++) and Celeb-DF (V2) datasets, achieving 97.99% (FF++) and 96.33% (CeDF) accuracy.

---

📂 Repository Structure

Assets/
│── Architecture.pdf        # Paper draft and architecture details
│── Pipeline.png            # Overall workflow of D4-Net
│── FF.png                  # Sample from FaceForensics++
│── CeDF.png                # Sample from Celeb-DF dataset
│── Error.png               # Misclassified sample illustration
│── CeDF_CM.png             # Confusion matrix for CeDF
│── FF++_CM.png             # Confusion matrix for FF++
│
├── dataloaders.py          # Dataset preparation and preprocessing
├── model.py                # D4-Net architecture and training loop
├── README.md               # Project documentation

---

🚀 Features

• Dual-path design combining spatial (RGB) and frequency (FFT) features.
• Cross-attention mechanism aligns modalities effectively.
• Adaptive feature fusion ensures robust decision-making.
• Lightweight yet high-performing (~26M params, 9 GFLOPs).
• Tested on benchmark datasets for generalization.

---

📊 Results

Dataset	Accuracy	F1 Score	AUC
FF++	97.99%	0.9800	0.9948
CeDF	96.33%	0.9725	0.9938

🔹 See detailed results and error analysis in Assets/ plots.

---

🛠 Installation

git clone https://github.com/omkarmakar/D4-Net.git
cd D4-Net

📑 Reference

If you use this work, please cite:

Om Karmakar, Sk Mohiuddin, Asfak Ali, Dmitrii Kaplun, Ram Sarkar.
D4-Net: Detecting Deepfakes using a Dual-branch Deep Learner.
Proceedings of the 33rd ACM International Conference on Multimedia (MM '25), Dublin, Ireland.
DOI : https://doi.org/10.1145/3746270.3760238

---

📬 Contact

For questions or collaborations:

• Om Karmakar – omkarmakar07@gmail.com
• Sk Mohiuddin – myselfmohiuddin@gmail.com