🧠 EfficientNet-B0 Deepfake Detector (FaceForensics++ C23)
👤 Author
Himanshu Kashyap (Xicor9)
MSc Advanced Computer Science
University of Strathclyde, Glasgow
📘 Overview
This repository contains a fine-tuned EfficientNet-B0 model for deepfake detection trained on the FaceForensics++ (FF++) C23 dataset. The model classifies face images or frames into:
The training data includes multiple manipulation types such as DeepFake, FaceSwap, Face2Face, and NeuralTextures. The model is designed for academic and research use.
📊 Performance
Frame-Level Metrics
| Metric | Score |
|---|
| AUC | 0.933 |
| AP | 0.898 |
| Accuracy | 0.852 |
| F1-Score | 0.843 |
Video-Level Metrics
(Mean probability across frames)
| Metric | Score |
|---|
| AUC | 0.94+ |
| Accuracy | 0.88+ |
🏗 Model Architecture & Training Setup
- Base model: EfficientNet-B0 (ImageNet pretrained)
- Final classification layer:
Linear(in_features, 2)
- Loss function: CrossEntropyLoss
- Optimizer: AdamW (lr = 3e-4, weight_decay = 1e-4)
- LR Scheduler: CosineAnnealingLR (T_max = 10)
- Mixed Precision (AMP): Enabled
- Epochs: 8
- Hardware: Tesla T4 (Google Colab)
🔧 How to Use This Model
Load the model
1import torch
2import torch.nn as nn
3from torchvision import models
4
5device = "cuda" if torch.cuda.is_available() else "cpu"
6
7# Load state dict from Hugging Face
8state_dict = torch.hub.load_state_dict_from_url(
9 "https://huggingface.co/Xicor9/efficientnet-b0-ffpp-c23/resolve/main/efficientnet_b0_ffpp_c23.pth",
10 map_location=device
11)
12
13# Rebuild architecture
14model = models.efficientnet_b0(weights=None)
15model.classifier[1] = nn.Linear(model.classifier[1].in_features, 2)
16
17model.load_state_dict(state_dict)
18model.to(device)
19model.eval()
📷 Example Inference
1from PIL import Image
2from torchvision import transforms
3
4transform = transforms.Compose([
5 transforms.Resize((224, 224)),
6 transforms.ToTensor()
7])
8
9img = Image.open("frame.jpg").convert("RGB")
10x = transform(img).unsqueeze(0).to(device)
11
12with torch.no_grad():
13 logits = model(x)
14 prob_fake = torch.softmax(logits, dim=1)[0][1].item()
15
16print("Fake Probability:", prob_fake)
🗄 Training Dataset
This model was trained on the FaceForensics++ (FF++) C23 dataset.
Included Video Types
- Real videos
- Manipulated videos, including:
- DeepFake
- FaceSwap
- Face2Face
- NeuralTextures
All videos were converted into frames and resized to 224 × 224 before training.
💡 Intended Use
This model is designed for:
- Deepfake research
- Academic coursework and university projects
- Benchmarking deepfake detection models
- Dataset experimentation and model comparison
❗ Not intended for real-world forensic, biometric, surveillance, or law-enforcement applications.
📎 License
This model is released strictly for research and educational purposes only.
Users must ensure compliance with dataset and model licenses before use.
❤️ Acknowledgements
- FaceForensics++ dataset authors
- PyTorch / TorchVision development teams
- EfficientNet authors
- Google Colab for compute resources
