Hugging Face's logo

Spaces:
MohsinElis
/
patch-mfft-api
Paused

App Files Community
patch-mfft-api / README.md
MohsinElis's picture
MohsinElis
Deploy production Patch-MFFT Tampering Forensics API
e83ff6c verified
preview code
|
Raw History Blame Contribute Delete
1.75 kB
metadata 
title: Patch-MFFT Tampering Forensics API
emoji: 🛡️
colorFrom: red
colorTo: indigo
sdk: docker
app_port: 7860
pinned: false
license: mit

Patch-MFFT: Dual-Branch Radial Frequency Transformer API

Production FastAPI forensics service engineered for localized generative inpainting, composite blackmail splicing, and style tampering detection.

Dual-Task Forensics Features

  1. 3-Way Holistic Forensic Triage:

    • [0] Authentic Camera Capture: Optical photon shot noise & Bayer demosaicing consistency
    • [1] Holistic Generative AI: Entire-frame deconvolution checkerboard harmonics
    • [2] Localized Spliced / Inpainted Tampering: Boundary phase incoherencies & local inpainting artifacts

  2. Pixel-Wise Tampering Localization:

    • Real-time dense anomaly probability mask $\hat{\mathbf{M}} \in [0, 1]^{384 \times 384}$ returned as a Base64-encoded PNG heatmap.

  3. Radial Annular Fourier Decomposition:

    • Concentric Low, Mid, and High frequency band energy distribution analysis.

Models Served

  • Patch-MFFT Tiny: 563K parameters (~4.2 ms GPU, 14.8 ms CPU, optimized for edge/mobile forensics)
  • Patch-MFFT Base: 3.28M parameters (97.84% accuracy, 81.00% mIoU localization score)
  • Patch-MFFT Large: 13.94M parameters (Maximum precision for subtle inpainting & blackmail composite detection)

API Endpoints

  • GET / — Healthcheck and status of loaded model variants
  • GET /models — Specifications and parameter counts of all variants
  • POST /predict?model=large — Multi-task forensic analysis on uploaded image
curl -X POST "https://mohsinelis-patch-mfft-api.hf.space/predict?model=large" \\
  -F "file=@sample.jpg"