🛡️ Detect The Deceptive — Multi-modal deepfake detection platform using ConvNeXt + FastAPI (in development)

⚠️ This project is actively under development. The architecture and roadmap below describe the planned system. Features are being built and validated incrementally. Contributions and feedback welcome.

A unified GenAI verification platform designed to detect AI-synthesized images and voice clones — addressing the growing threat of deepfakes in misinformation, identity fraud, and social engineering attacks.

---

🎯 Problem Statement

Generative AI has made it trivially easy to fabricate realistic faces, voices, and media. Existing detection tools are siloed — separate tools for image forgery, separate tools for audio cloning. Detect The Deceptive aims to provide a single API and dashboard that handles both modalities, with explainable outputs that go beyond a binary "fake/real" verdict.

---

🏗️ Planned Architecture

User Upload (Image / Audio)
          │
          ▼
┌────────────────────────────────┐
│        React Frontend          │
│  Upload → Results → Visual     │
└──────────────┬─────────────────┘
               │ REST API
               ▼
┌────────────────────────────────┐
│    FastAPI Microservices       │
│  ┌────────────┐ ┌────────────┐ │
│  │ /analyze/  │ │ /analyze/  │ │
│  │   image    │ │   voice    │ │
│  └──────┬─────┘ └──────┬─────┘ │
└─────────┼──────────────┼───────┘
          │              │
          ▼              ▼
  ConvNeXt Model   Audio Classifier
  + Grad-CAM        + Feature
  Explainability      Analysis
          │              │
          ▼              ▼
     Metadata       Confidence
     Forensics        Score
          │              │
          └──────┬────────┘
                 ▼
          Unified Verdict
          + Explanation

---

🔬 Technical Approach

Image Deepfake Detection

• Model: ConvNeXt-Tiny (chosen for strong performance on texture artifacts vs. ResNet/EfficientNet)
• Training data: FaceForensics++, DFDC datasets (planned)
• Explainability: Grad-CAM activation maps to highlight regions that triggered the detection
• Metadata forensics: EXIF analysis, DCT coefficient analysis, noise pattern inconsistency

Voice Clone Detection

• Feature extraction: MFCC, spectral centroid, zero-crossing rate analysis
• Detection of unnatural prosody patterns and frequency artifacts common in TTS/voice conversion
• Classifier: SVM or lightweight CNN over audio spectrograms (under evaluation)

---

📍 Current Development Status

Component	Status
Project architecture design	✅ Complete
FastAPI backend scaffolding	🔄 In Progress
ConvNeXt model training pipeline	🔄 In Progress
Grad-CAM visualization	🔄 In Progress
Voice analysis module	📋 Planned
React frontend	📋 Planned
Docker containerization	📋 Planned
Integration tests	📋 Planned

---

🚀 Getting Started (Development Build)

git clone https://github.com/MeghnaaT/Detect-The-Deceptive.git
cd Detect-The-Deceptive

# Backend
cd backend
pip install -r requirements.txt
uvicorn main:app --reload

Full setup instructions will be updated as each module reaches a runnable state.

---

📁 Repository Structure (Planned)

├── backend/
│   ├── main.py                  # FastAPI entry point
│   ├── routers/
│   │   ├── image_analysis.py    # ConvNeXt inference + Grad-CAM
│   │   └── voice_analysis.py    # Audio classifier
│   ├── models/                  # Saved model weights
│   └── utils/
│       ├── metadata_forensics.py
│       └── gradcam.py
├── frontend/                    # React app (planned)
├── docker-compose.yml
├── requirements.txt
└── README.md

---

🔮 Roadmap

• Complete ConvNeXt training pipeline with benchmark results on FaceForensics++
• Implement Grad-CAM heatmap overlay in API response
• Build and validate voice clone detection module
• Integrate both into unified FastAPI service
• Build React dashboard with upload + results visualization
• Dockerize and deploy to cloud (Hugging Face Spaces / Render)
• Publish model card with accuracy metrics and dataset details

---

🤝 Contributing

This project is in early development — contributions, ideas, and issue reports are especially welcome at this stage.

1. Fork the repository
2. Create a branch: git checkout -b feat/your-contribution
3. Open a PR or issue with a description of what you're working on

---

📄 License

MIT License — see LICENSE for details.

---

Built by Meghna Tiwari · Deepfake detection research & tooling