FaceForge Detector: State-of-the-Art Deepfake Detection

🎯 99.33% Accuracy | 0.9995 AUC-ROC | 0.67% Error Rate

Model Description

FaceForge Detector is a high-performance deepfake detection model based on XceptionNet architecture that achieves state-of-the-art results on the FaceForensics++ dataset. This model can distinguish authentic faces from AI-generated deepfakes with exceptional accuracy.

Key Features:

🚀 99.33% accuracy on test set (1,500 samples)
📊 0.9995 AUC-ROC score
⚡ <200ms inference time per image
🎯 Only 2 false negatives out of 750 deepfakes (0.27% miss rate)
🔧 22M trainable parameters
💻 Trained efficiently on CPU (7.5 hours)

Model Architecture

XceptionNet Backbone (20.8M params)
├── Entry Flow
├── Middle Flow (8 blocks)
├── Exit Flow
├── Global Average Pooling
└── Custom Classification Head (1.1M params)
    ├── Dropout (p=0.5)
    ├── FC(2048 → 512) + ReLU
    ├── Dropout (p=0.3)
    └── FC(512 → 2) [Real/Fake]

Performance Metrics

Test Set Results (1,500 samples)

Metric	Value
Accuracy	99.33%
AUC-ROC	0.9995
Precision	98.94%
Recall	99.73%
F1 Score	0.9934
Specificity	98.93%

Confusion Matrix

	Predicted Real	Predicted Fake
Actual Real	742 (TN)	8 (FP)
Actual Fake	2 (FN)	748 (TP)

Total Errors: 10 out of 1,500 (0.67%)

Comparison with State-of-the-Art

Method	Accuracy	AUC-ROC
MesoNet	83.1%	0.847
XceptionNet (base)	95.3%	0.969
EfficientNet-B4	96.8%	0.981
Capsule Network	97.4%	0.986
Face X-ray	98.1%	0.991
FaceForge (Ours)	99.33%	0.9995

Usage

Installation

pip install torch torchvision timm pillow numpy

Inference

import torch
import timm
from PIL import Image
from torchvision import transforms

# Load model
model = timm.create_model('xception', pretrained=False, num_classes=2)
checkpoint = torch.load('detector_best.pth', map_location='cpu')
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Preprocessing
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
])

# Inference
def detect_deepfake(image_path):
    image = Image.open(image_path).convert('RGB')
    input_tensor = transform(image).unsqueeze(0)
    
    with torch.no_grad():
        logits = model(input_tensor)
        probs = torch.softmax(logits, dim=1)
        prediction = torch.argmax(probs, dim=1).item()
        confidence = probs[0][prediction].item()
    
    label = "REAL" if prediction == 0 else "FAKE"
    return label, confidence

# Example
label, confidence = detect_deepfake("face.jpg")
print(f"Prediction: {label} (Confidence: {confidence:.2%})")

Gradio Demo

import gradio as gr

def predict(image):
    label, confidence = detect_deepfake(image)
    return f"{label} ({confidence:.2%})"

demo = gr.Interface(
    fn=predict,
    inputs=gr.Image(type="filepath"),
    outputs="text",
    title="FaceForge Deepfake Detector",
    description="Upload a face image to detect if it's real or AI-generated"
)

demo.launch()

Training Details

Dataset

Source: FaceForensics++ (c40 compression)
Training: 7,000 images (3,500 real + 3,500 fake)
Validation: 1,500 images (750 real + 750 fake)
Test: 1,500 images (750 real + 750 fake)

Hyperparameters

optimizer: AdamW
learning_rate: 1e-4
weight_decay: 1e-4
batch_size: 32
epochs: 10
lr_schedule: Cosine Annealing (1e-4 → 1e-6)
augmentation:
  - Random Horizontal Flip (p=0.5)
  - Color Jitter (brightness=0.2, contrast=0.2, saturation=0.2)

Training Time

Total: 7.52 hours (451.5 minutes)
Per Epoch: ~45 minutes
Hardware: CPU (no GPU required)
Best Checkpoint: Epoch 8 (0.9998 AUC-ROC)

Limitations

Dataset Scope: Trained on FaceForensics++ deepfakes; may need fine-tuning for other manipulation methods
Single Frame: Processes individual images; doesn't leverage temporal information from videos
Compression: Trained on c40 compression; performance may vary with different quality levels
Domain: Optimized for face-centric images; may struggle with partial faces or unusual angles

Ethical Considerations

This model is intended for: ✅ Research and education ✅ Content moderation ✅ Forensic analysis ✅ Fact-checking

Not intended for: ❌ Malicious surveillance ❌ Discriminatory profiling ❌ Invasion of privacy

Citation

@techreport{nasir2026faceforge,
  title={FaceForge: A Deep Learning Framework for Facial Manipulation Generation and Detection},
  author={Nasir, Huzaifa},
  institution={National University of Computer and Emerging Sciences},
  year={2026},
  doi={10.5281/zenodo.18530439}
}

License

This model is released under CC BY 4.0 license. See LICENSE file for details.

Author

Huzaifa Nasir
National University of Computer and Emerging Sciences (NUCES)
Islamabad, Pakistan
📧 nasirhuzaifa95@gmail.com

Acknowledgments

FaceForensics++ dataset creators
PyTorch and timm library developers
XceptionNet architecture (François Chollet)

Downloads last month: -; Downloads are not tracked for this model. How to track

huzaifanasirrr
/

faceforge-detector