ash12321/deepfake-detector-v15

🧠 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

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