A full-stack, cloud-deployed machine learning application that detects synthetic anomalies and deepfakes. Rather than relying solely on spatial pixel data, this engine transforms media into the frequency domain to uncover microscopic digital artifacts left behind by GANs and modern Diffusion models.
Auto-generated Swagger UI for backend API testing
Standard convolutional neural networks (CNNs) often struggle against modern high-fidelity diffusion models (like Midjourney or Flux) when looking only at spatial pixels. To solve this, this pipeline utilizes a 2D Fast Fourier Transform (FFT) preprocessing step.
The Pipeline:
- Micro-Blurring (3x3 Kernel): A delicate Gaussian blur filters out standard camera sensor noise and Moire patterns without erasing the high-frequency synthetic artifacts.
- Frequency Domain Transformation: The image is converted to grayscale and passed through a 2D FFT to generate a magnitude spectrum.
- Anomaly Classification: The resulting frequency map is fed into a custom-tuned ResNet-18 architecture. Because AI generators leave periodic, grid-like spectral fingerprints in the frequency domain, the ResNet easily identifies synthetic media that looks flawless to the human eye.
Backend (Inference Engine) - Hosted on Hugging Face Spaces (Docker)
- PyTorch & Torchvision: Custom ResNet-18 model loading and tensor operations.
- OpenCV (cv2) & NumPy: Image processing, matrix manipulation, and FFT math.
- FastAPI & Uvicorn: High-performance, asynchronous REST API. Provides auto-generated OpenAPI (Swagger) documentation.
Frontend (Client UI) - Hosted on Vercel
- React & Vite: Fast, responsive user interface.
- Tailwind CSS: Modern, forensic-style dark mode UI with interactive Bento-grid results.
- Axios: Handles multipart/form-data tensor transmission to the cloud backend.
Try the live engine here: deepfake-detection-engine.vercel.app
Note: The backend inference engine is hosted on a free Hugging Face Docker Space. If it hasn't been used in 48 hours, it may take 1-2 minutes to "wake up" on your first scan.
If you wish to run the engine locally for development:
git clone [https://github.com/anushkadas/deepfake-detection-engine.git](https://github.com/anushkadas/deepfake-detection-engine.git)
cd deepfake-detection-engine