File size: 3,180 Bytes
c9721cb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 | # 🧠 DeepFake Detector V15
**Self-Learning Deepfake Detector with Web Search Integration**
## ✨ Features
- 🔍 **Real Web Search** - SerpAPI reverse image + Serper text search
- 🧠 **Self-Learning** - Improves from user feedback
- 🛡️ **EWC Protection** - Never forgets old knowledge
- 📈 **Progressive** - Gets smarter over time
## 📊 Architecture
| Component | Parameters | Trainable |
|-----------|------------|-----------|
| Swin-Large Backbone | 197M | ❌ Frozen |
| Adapter Layers | 1.5M | ✅ Yes |
| **Total** | **198.5M** | 1.5M |
## 🚀 Quick Start
```python
import torch
import timm
from safetensors.torch import load_file
from torchvision import transforms
from PIL import Image
class DeepfakeDetector(torch.nn.Module):
def __init__(self):
super().__init__()
self.backbone = timm.create_model('swin_large_patch4_window7_224',
pretrained=False, num_classes=0)
feat_dim = 1536
self.adapter = torch.nn.Sequential(
torch.nn.Linear(feat_dim, 512),
torch.nn.LayerNorm(512),
torch.nn.ReLU(),
torch.nn.Dropout(0.1),
torch.nn.Linear(512, feat_dim)
)
self.classifier = torch.nn.Sequential(
torch.nn.Linear(feat_dim, 512),
torch.nn.BatchNorm1d(512),
torch.nn.GELU(),
torch.nn.Dropout(0.3),
torch.nn.Linear(512, 128),
torch.nn.BatchNorm1d(128),
torch.nn.GELU(),
torch.nn.Dropout(0.15),
torch.nn.Linear(128, 1)
)
def forward(self, x):
features = self.backbone(x)
adapted = features + 0.1 * self.adapter(features)
return self.classifier(adapted).squeeze(-1)
# Load
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = DeepfakeDetector()
model.load_state_dict(load_file("model.safetensors"))
model = model.to(device)
model.eval()
# Preprocess
transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
# Predict
image = Image.open("test.jpg").convert("RGB")
with torch.no_grad():
prob = torch.sigmoid(model(transform(image).unsqueeze(0).to(device))).item()
print(f"Fake: {prob:.1%}" if prob > 0.5 else f"Real: {1-prob:.1%}")
```
## 📈 Performance
| Version | F1 Score | Improvement |
|---------|----------|-------------|
| V14 Base | 0.9586 | - |
| V15 (+50 samples) | ~0.962 | +0.3% |
| V15 (+200 samples) | ~0.968 | +1.0% |
| V15 (+500 samples) | ~0.975 | +1.6% |
## 🌐 Web Search Integration
V15 uses two APIs for verification:
- **SerpAPI** - Google reverse image search (finds where image exists online)
- **Serper.dev** - Text search (finds deepfake mentions)
## 🔧 Self-Learning
Uses **Elastic Weight Consolidation (EWC)** to:
- Learn from new user feedback
- Without forgetting previous knowledge
- Only trains adapter layers (fast!)
## 📚 Model Lineage
`V12 → V13 → V14 → V15 (Self-Learning)`
## 📄 License
MIT
---
**Built with PyTorch, timm, and Gradio**
|