🚀 WYT-Net: Wavelet-YOLO-Transformer Network

Lightweight Hybrid YOLOv8 Framework for Real-Time Deepfake Detection on Edge Devices

---

📖 Overview

WYT-Net (Wavelet-YOLO-Transformer Network) is a lightweight, frequency-aware deep learning architecture designed for real-time deepfake detection on edge devices.

Unlike heavy Vision Transformers or deep CNN-based forensic models, WYT-Net:

• Integrates Discrete Wavelet Transform (DWT) for frequency-domain artifact detection.
• Uses SimAM parameter-free attention.
• Incorporates FastViT-inspired hybrid CNN-Transformer blocks.
• Maintains a highly efficient footprint of < 2.6 Million parameters.
• Runs efficiently on edge hardware like Raspberry Pi 4 and NVIDIA Jetson Nano.

The framework achieves an optimal balance between accuracy, efficiency, and deployability.

---

⚠️ Problem Statement

The rapid rise of high-fidelity deepfake media threatens digital integrity, journalism, biometric systems, and public trust.

Existing Limitations:

• Heavy Models: Architectures like ResNet, ViT, and EfficientNet are computationally expensive.
• Hardware Dependency: High reliance on server-grade GPUs for inference.
• Edge Unfriendly: Poor feasibility for deployment on mobile or IoT devices.
• Spatial Bias: Limited frequency-domain awareness allows subtle generative artifacts to go unnoticed.

Key Insight:

GANs and Diffusion models leave high-frequency artifacts that spatial CNNs struggle to capture. WYT-Net addresses this using wavelet-domain feature engineering combined with a lightweight hybrid architecture.

---

🎯 Objectives

• ✅ Frequency-Domain Artifact Detection: Achieve this via Wavelet Decomposition.
• ✅ Architectural Efficiency: Design a hybrid CNN-Transformer backbone under 2.6M parameters.
• ✅ High Accuracy: Surpass 94% classification accuracy on challenging datasets.
• ✅ Edge Deployment: Guarantee real-time performance on edge devices.
• ✅ Robust Generalization: Ensure resilience against real-world artifacts using comprehensive data augmentation.

---

🛠️ Methodology

1️⃣ Feature Engineering – 4-Channel Wavelet Input

Instead of standard RGB, WYT-Net uses Daubechies-2 (db2) Wavelet Decomposition:

Channel	Description
LL	Approximation (Global Spatial Features)
LH	Horizontal Details (High-Frequency)
HL	Vertical Details (High-Frequency)
HH	Diagonal Details (High-Frequency)

This explicit separation exposes forgery-induced noise patterns that are otherwise invisible in the RGB spectrum.

2️⃣ Proposed Architecture – WYT-Net

Modified from the YOLOv8n backbone using architectural "surgery":

• Split Backbone ➡️ Multi-scale feature extraction.
• SimAM Attention ➡️ Parameter-free 3D attention refinement for localized texture traces.
• FastViT Block ➡️ Structural reparameterization + token mixing for global structural consistency.
• Feature Fusion ➡️ Dual-stream Global Average Pooling (GAP) concatenation.
• Final MLP Head ➡️ Optimized binary classification.

---

📊 Experimental Results

🔹 Performance Comparison

Metric	Baseline YOLOv8n	WYT-Net (Proposed)
Accuracy	91.45%	94.44%
mAP@0.5	0.8521	0.9371
F1-Score (Fake)	0.7145	0.8022
Parameters	2.04 M	2.57 M
GFLOPs	0.2498	0.2841

🔹 Detailed Classification Report (WYT-Net)

Class	Precision	Recall	F1-Score	Support
Real	0.9735	0.9619	0.9677	23,834
Fake	0.7738	0.8328	0.8022	3,731
Average / Total	0.9465	0.9444	0.9453	27,565

---

🚀 Edge Deployment Results

🔹 Raspberry Pi 4 (NCNN - FP16)

• Inference Engine: Tencent NCNN
• Latency: ~124.5 ms per frame
• FPS: ~8 FPS
• Optimization: Multi-threading (4 cores) + ARM Neon SIMD

🔹 NVIDIA Jetson Nano (TensorRT)

Precision	Latency
FP32	28.5 ms
FP16	14.2 ms
INT8	8.1 ms

• Peak Power Consumption: ~3.1W
• Performance: Near real-time inference achieved natively on edge hardware.

---

📂 Repository Structure

.
├── Models/
│   ├── yolov8_ai_augmented.py   # Main WYT-Net implementation
│   ├── yolov8_hybrid_ai.py      # Hybrid backbone variant
│   ├── augment_dataset.py       # Data augmentation pipeline
│   └── extract_faces.py         # Face detection & DWT preprocessing
├── Results/
│   ├── results_journal_Yolov8_Ai_augmented/ # Proposed model metrics & plots
│   ├── results_hybrid_no_ai_augmented/      # Ablation study metrics
│   └── results_baseline/                    # Reference benchmark metrics
├── test_model.py                # Inference and validation script
└── requirements.txt             # Project dependencies

⚡ Quick Start

🔹 Installation

git clone [https://github.com/codewithyug06/deepfake-detection-pipelines.git](https://github.com/codewithyug06/deepfake-detection-pipelines.git)
cd deepfake-detection-pipelines
pip install -r requirements.txt

---

📦 Dataset

The models were trained and evaluated on data derived from the Celeb-DF (v2) dataset.

• Original Dataset: 27,565 images
• Augmented Dataset: 3× expansion resulting in a more balanced and robust training set

🔹 Augmentations Applied

• Random Rotation (±15°)
• Horizontal Flip
• Color Jittering

---

🔬 Ablation Study

Configuration	Accuracy
Baseline YOLOv8n	91.45%
+ Wavelet Input	92.81%
+ SimAM Attention	93.56%
+ FastViT (WYT-Net)	94.44%

Key Takeaway:
Wavelet-domain information, when combined with hybrid global reasoning via FastViT, significantly improves deepfake detection reliability.

---

🧠 Key Contributions

• First of Its Kind: Lightweight YOLO-based frequency-aware deepfake detector explicitly optimized for edge computing
• Extreme Efficiency: Highly pruned hybrid CNN-Transformer architecture with fewer than 2.6M parameters
• Edge-Ready: Proven real-time inference on Raspberry Pi 4 and NVIDIA Jetson Nano
• Optimal Tradeoff: Superior balance between detection accuracy and computational complexity compared to state-of-the-art forensic models

---