A sleek web app that flags AI-generated video in seconds with frame-by-frame analysis.

Built for ACMHacks 2025 — UCSC ACM’s annual hackathon.
Team: Sayan Sharma, Yash Khot.

Upload a short video (.mp4/.mov). DeepSight:

• Detects faces per frame (OpenCV)
• Runs a fine-tuned Xception model (FaceForensics++ weights)
• Provides a deepfake probability score and a synced frame-by-frame chart
• Returns an overall verdict + JSON for further analysis

Frontend

• React (Vite)
• Modern, dark UI (CSS)
• Video panel + probability HUD
• Live chart rendering (frame vs. fake probability) (MatPlotLib)

Backend

• FastAPI + Uvicorn
• Multipart uploads
• Model inference endpoint /api/predict

ML / CV

• PyTorch, torchvision, timm (Xception backbone)
• OpenCV (face detection)
• NumPy, Pandas, Matplotlib (frame analytics/plots)
• MoviePy (video IO)

1. Build the Docker Image

• docker build -t deepsight .

2. Run the Container

• docker run -p 8000:8000 deepsight

3. Open DeepSight

• Visit → http://localhost:8000

Requirements: Python 3.10–3.12, Node 18+, ffmpeg installed.

1. Clone

• git clone https://github.com/Sayan090807/DeepSight.git
• cd DeepSight

2. Python deps

• python -m venv venv && source venv/bin/activate # Windows: venv\Scripts\activate
• pip install -r requirements.txt

3. Pull LFS weights

• git lfs install
• git lfs pull # downloads the FaceForensics++ Xception weights

4. Build the React UI

• cd app/frontend
• npm install
• npm run build
• cd ../../

5. Run Project

• make run
• Open http://0.0.0.0:8000

make build # builds the React frontend make run # starts FastAPI (serves the built UI) make clean # removes frontend build artifacts

Multipart form upload (file).

Response Example: { "frames_analyzed": 30, "average_fake_probability": 0.12, "frame_probs": [0.08, 0.11, 0.14, "..."], "plot_png": "/plots/last_run.png" }

• Model: Xception trained on FaceForensics++
• Safe loading: Weights from older checkpoints are handled defensively
• Performance: Frame skipping + face cropping → real-time on CPU
• Explainability: Per-frame timeline exposes where the model spikes

Skills: ML inference, computer vision, web dev, API design, data visualization, model integration, deployment.

Libraries:
FastAPI, Uvicorn, PyTorch, torchvision, timm, OpenCV, MoviePy, NumPy, Pandas, Matplotlib, React (Vite).

• FaceForensics++ and the broader deepfake research community
• FastAPI & PyTorch teams
• UCSC ACM organizers and mentors for ACMHacks 2025

MIT — see LICENSE.