Spaces:

MohsinElis
/

mfft-api

Paused

App Files Files Community

mfft-api / README.md

MohsinElis

Deploy production MFFT detection API with verified 20-epoch weights

37b0e8b verified 26 days ago

preview code

Raw History Blame Contribute Delete

1.38 kB

---
title: MFFT AI Image Detector API
emoji: 🛡️
colorFrom: indigo
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
license: mit
---

# MFFT: Multi-Frequency Fusion Transformer — Production Detection API

FastAPI inference server for AI-generated image and deepfake detection based on radial Fourier frequency sub-band decomposition and Frequency-Guided Attention (FGA).

Three verified production models trained for 20 epochs on 120,000 images are served concurrently:
- **MFFT-Tiny** (372K params, 1.49 MB FP16): **98.59% accuracy**, **99.88% AUC-ROC** (~3.2 ms latency)
- **MFFT-Base** (1.62M params, 6.49 MB FP16): **98.37% accuracy**, **99.87% AUC-ROC** (5-seed replication average: **98.58% ± 0.13%**)
- **MFFT-Large** (6.30M params, 25.21 MB FP16): **98.11% accuracy**, **99.68% AUC-ROC** (Default production profile)

## Endpoints

- `GET /` — Healthcheck and loaded model status
- `GET /models` — Available model profiles and specifications
- `POST /predict?model=large` — Single image forensic analysis with anomaly heatmap (base64 PNG) and radial Fourier frequency band breakdown
- `POST /predict/batch?model=large` — Multi-image batch inference

## Example Usage

```bash
curl -X POST "https://mohsinelis-mfft-api.hf.space/predict?model=large" \
  -H "Authorization: Bearer free" \
  -F "file=@image.jpg"
```

	---
	title: MFFT AI Image Detector API
	emoji: 🛡️
	colorFrom: indigo
	colorTo: blue
	sdk: docker
	app_port: 7860
	pinned: false
	license: mit
	---

	# MFFT: Multi-Frequency Fusion Transformer — Production Detection API

	FastAPI inference server for AI-generated image and deepfake detection based on radial Fourier frequency sub-band decomposition and Frequency-Guided Attention (FGA).

	Three verified production models trained for 20 epochs on 120,000 images are served concurrently:
	- MFFT-Tiny (372K params, 1.49 MB FP16): 98.59% accuracy, 99.88% AUC-ROC (~3.2 ms latency)
	- MFFT-Base (1.62M params, 6.49 MB FP16): 98.37% accuracy, 99.87% AUC-ROC (5-seed replication average: 98.58% ± 0.13%)
	- MFFT-Large (6.30M params, 25.21 MB FP16): 98.11% accuracy, 99.68% AUC-ROC (Default production profile)

	## Endpoints

	- `GET /` — Healthcheck and loaded model status
	- `GET /models` — Available model profiles and specifications
	- `POST /predict?model=large` — Single image forensic analysis with anomaly heatmap (base64 PNG) and radial Fourier frequency band breakdown
	- `POST /predict/batch?model=large` — Multi-image batch inference

	## Example Usage

	```bash
	curl -X POST "https://mohsinelis-mfft-api.hf.space/predict?model=large" \
	-H "Authorization: Bearer free" \
	-F "file=@image.jpg"
	```