SnapSure

SnapSure is a local deepfake detection app with:

• frontend/: Next.js frontend on port 3000
• backend/: Flask API on port 8000
• models/: inference code for the ensemble detector
• docker/: backend and frontend Dockerfiles
• k8s/: Kubernetes manifests for Minikube
• Jenkinsfile: Jenkins pipeline for local Docker and Minikube deployment

What it does

The backend loads two Hugging Face image-classification models:

• Wvolf/ViT_Deepfake_Detection
• dima806/deepfake_vs_real_image_detection

It averages the fake probability from both models and returns:

• overall_label
• overall_confidence
• fake_score
• num_faces

Face detection uses MTCNN. Face count is returned in the response, but face detection is not used to make the REAL/FAKE decision.

API

GET /health

Returns:

{ "status": "ok", "model": "ensemble" }

If DEMO_MODE=true, it returns:

{ "status": "ok", "model": "demo" }

POST /predict

Send multipart/form-data with a file field.

Supported file types:

• .jpg
• .jpeg
• .png
• .webp

Example success response:

{
  "overall_label": "REAL",
  "overall_confidence": 0.8765,
  "fake_score": 0.1235,
  "num_faces": 1,
  "faces": []
}

Run with Docker Compose

From the repo root:

docker compose up --build

App URLs:

• Frontend: http://localhost:3000
• Backend: http://localhost:8000

Stop it with:

docker compose down

Model cache

The backend container stores Hugging Face and Torch cache in a named Docker volume:

• backend-model-cache

This means:

• the models are downloaded on the first run
• later restarts reuse the cache
• the model weights are not baked into the backend image

Run locally without Docker

Backend:

cd backend
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
cd ..
set PYTHONPATH=.
python -m backend.app

Frontend:

cd frontend
npm install
npm run dev

Environment

Backend environment values are read from:

• backend/.env
• backend/.env.docker

Main backend variables:

• MODEL_DEVICE=cpu
• DEMO_MODE=false

The Docker and Kubernetes setups also set:

• HF_HOME
• TORCH_HOME
• XDG_CACHE_HOME

These keep downloaded model files in a persistent cache path.

Docker image note

The backend image now uses CPU-only PyTorch wheels and excludes local virtualenv folders from the build context. That keeps the backend image much smaller than the earlier CUDA-heavy build.

Kubernetes and Jenkins

For Minikube deployment:

• see k8s/README-K8S.md

For Jenkins setup:

• see JENKINS-SETUP.md