# Deepfake Detector V8 - Foundation Model

**Baseline accuracy: 100.00%** | **F1-Score: 1.0000**

This is a foundation model designed for continuous improvement. It provides a solid starting point with room for enhancement.

## 📊 Performance

- **Accuracy**: 100.00%
- **F1-Score**: 1.0000
- **Precision**: 1.0000
- **Recall**: 1.0000
- **Training Samples**: 4,800
- **Validation Samples**: 1,200

## 🏗️ Architecture

- **Backbone**: EfficientNetV2-S (pretrained on ImageNet)
- **Classifier**: 3-layer MLP with BatchNorm and Dropout
- **Input Size**: 224x224 RGB images
- **Output**: Binary classification (Real vs Fake)

## 📁 Files in This Folder

```
./deepfake_v8_model/
├── model.safetensors       ✓ Model weights (HuggingFace format)
├── pytorch_model.bin       ✓ Fallback weights (if safetensors unavailable)
├── optimizer.pt            ✓ Optimizer state for continuing training
├── scheduler.pt            ✓ Learning rate scheduler state
├── config.json             ✓ Model architecture configuration
├── training_args.json      ✓ All hyperparameters used
├── rng_state.pth          ✓ Random states for reproducibility
├── metrics.json           ✓ Performance metrics
├── tokenizer.json         ✓ HuggingFace compatibility
├── tokenizer_config.json  ✓ HuggingFace compatibility
└── README.md              ✓ This file
```

## 🚀 How to Use This Model

### Option 1: Inference (Predict on New Images)

```python
import torch
import torch.nn as nn
import timm
from PIL import Image
from torchvision import transforms

# Define model architecture
class DeepfakeDetector(nn.Module):
    def __init__(self, dropout=0.5):
        super().__init__()
        self.backbone = timm.create_model('tf_efficientnetv2_s', pretrained=False, num_classes=0)
        self.classifier = nn.Sequential(
            nn.Linear(1280, 512), nn.BatchNorm1d(512), nn.SiLU(), nn.Dropout(dropout),
            nn.Linear(512, 256), nn.BatchNorm1d(256), nn.SiLU(), nn.Dropout(dropout * 0.8),
            nn.Linear(256, 1)
        )
    
    def forward(self, x):
        return self.classifier(self.backbone(x)).squeeze(-1)

# Load model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = DeepfakeDetector(0.5)

try:
    from safetensors.torch import load_file
    state_dict = load_file('./deepfake_v8_model/model.safetensors')
except:
    state_dict = torch.load('./deepfake_v8_model/pytorch_model.bin', map_location=device)

model.load_state_dict(state_dict)
model = model.to(device)
model.eval()

# Predict on image
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])

img = Image.open('test_image.jpg').convert('RGB')
img_tensor = transform(img).unsqueeze(0).to(device)

with torch.no_grad():
    logit = model(img_tensor)
    prob = torch.sigmoid(logit).item()

if prob > 0.7:
    print(f"🔴 LIKELY FAKE ({prob:.1%} confidence)")
elif prob > 0.5:
    print(f"⚠️ POSSIBLY FAKE ({prob:.1%} confidence)")
elif prob > 0.3:
    print(f"⚠️ POSSIBLY REAL ({(1-prob):.1%} confidence)")
else:
    print(f"🟢 LIKELY REAL ({(1-prob):.1%} confidence)")
```

### Option 2: Continue Training (Improve the Model)

```python
import torch
import torch.nn as nn
import torch.optim as optim
import timm
import json
import random
import numpy as np

# Load training configuration
with open('./deepfake_v8_model/training_args.json', 'r') as f:
    config = json.load(f)

# Define model (same architecture as above)
class DeepfakeDetector(nn.Module):
    def __init__(self, dropout=0.5):
        super().__init__()
        self.backbone = timm.create_model('tf_efficientnetv2_s', pretrained=False, num_classes=0)
        self.classifier = nn.Sequential(
            nn.Linear(1280, 512), nn.BatchNorm1d(512), nn.SiLU(), nn.Dropout(dropout),
            nn.Linear(512, 256), nn.BatchNorm1d(256), nn.SiLU(), nn.Dropout(dropout * 0.8),
            nn.Linear(256, 1)
        )
    
    def forward(self, x):
        return self.classifier(self.backbone(x)).squeeze(-1)

# Load model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = DeepfakeDetector(config['dropout'])

try:
    from safetensors.torch import load_file
    state_dict = load_file('./deepfake_v8_model/model.safetensors')
except:
    state_dict = torch.load('./deepfake_v8_model/pytorch_model.bin', map_location=device)

model.load_state_dict(state_dict)
model = model.to(device)

# Load optimizer
optimizer = optim.AdamW(
    filter(lambda p: p.requires_grad, model.parameters()),
    lr=config['learning_rate'] * 0.1  # Lower LR for fine-tuning
)
optimizer.load_state_dict(torch.load('./deepfake_v8_model/optimizer.pt'))

# Load scheduler
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=config['epochs'])
scheduler.load_state_dict(torch.load('./deepfake_v8_model/scheduler.pt'))

# Load RNG state for reproducibility
rng_state = torch.load('./deepfake_v8_model/rng_state.pth')
random.setstate(rng_state['python'])
np.random.set_state(rng_state['numpy'])
torch.set_rng_state(rng_state['torch'])
if torch.cuda.is_available() and rng_state['torch_cuda']:
    torch.cuda.set_rng_state_all(rng_state['torch_cuda'])

print("✓ Model loaded and ready for continued training!")

# Now continue training with:
# - More epochs
# - More/better data
# - Different augmentations
# - Fine-tuning strategies
```

## 📤 How to Upload to HuggingFace

### Step 1: Install HuggingFace CLI

```bash
pip install -U huggingface_hub
```

### Step 2: Login to HuggingFace

```python
from huggingface_hub import notebook_login
notebook_login()
```

Or via CLI:
```bash
huggingface-cli login
```

### Step 3: Upload Your Model

```python
from huggingface_hub import HfApi

api = HfApi()
api.upload_folder(
    folder_path="./deepfake_v8_model",
    repo_id="YOUR_USERNAME/deepfake-detector-v8",
    repo_type="model"
)

print("✓ Model uploaded to HuggingFace!")
print("Visit: https://huggingface.co/YOUR_USERNAME/deepfake-detector-v8")
```

### Step 4: Load from HuggingFace in New Notebook

```python
from huggingface_hub import hf_hub_download
import torch

# Download files
model_path = hf_hub_download(
    repo_id="YOUR_USERNAME/deepfake-detector-v8",
    filename="model.safetensors"
)

optimizer_path = hf_hub_download(
    repo_id="YOUR_USERNAME/deepfake-detector-v8",
    filename="optimizer.pt"
)

# Load and continue training...
```

## 💡 How to Improve This Model

### 1. Add More Real Data
```python
# In new notebook, load more datasets
from datasets import load_dataset

# Add real AI-generated images
stable_diffusion_images = load_dataset("your_sd_dataset")
dalle_images = load_dataset("your_dalle_dataset")

# Combine with existing model and retrain
```

### 2. Train Longer
```python
# Load model (as shown above)
# Then train for 5-10 more epochs
CONFIG['epochs'] = 10  # or more
# Continue training loop...
```

### 3. Unfreeze More Layers
```python
# Unfreeze all backbone layers for fine-tuning
for param in model.backbone.parameters():
    param.requires_grad = True

# Use lower learning rate
optimizer = optim.AdamW(model.parameters(), lr=1e-5)
```

### 4. Use Real Deepfakes
```python
# Load actual deepfake datasets
# - FaceForensics++
# - Celeb-DF
# - DFDC
# And retrain on real deepfakes
```

### 5. Ensemble Multiple Models
```python
# Train multiple versions
# Average predictions for better accuracy
```

## 🎯 Improvement Roadmap

**Current**: 100.0% (Foundation)

**Next Steps**:
1. **Add 10K more samples** → Target: 93-95%
2. **Train 5 more epochs** → Target: 94-96%
3. **Add real AI images** → Target: 95-97%
4. **Fine-tune all layers** → Target: 96-98%
5. **Ensemble 3 models** → Target: 97-99%

## ⚠️ Important Notes

- This model was trained on **synthetic fakes**
- For production use, train on **real AI-generated images**
- Use **confidence thresholds** (>0.7 for high confidence)
- Always **validate on diverse test sets**
- Consider **ethical implications** of deployment

## 📝 License

MIT License - Free to use with attribution

## 🤝 Contributing

This is a foundation model meant to be improved! Feel free to:
- Add more training data
- Experiment with architectures
- Share your improvements
- Create better versions

## 📧 Contact

For questions or improvements, open an issue on the repository.

---

**Generated by**: Deepfake Detector V8
**Training Time**: 35.1 minutes
**Date**: 2025-10-22 20:27:12