metadata

title: MCP Hackathon Deepfake Watchdog
emoji: 🛡️
colorFrom: green
colorTo: gray
sdk: gradio
sdk_version: 5.33.1
app_file: app.py
pinned: false
license: mit
short_description: Upload your image and/or voice to scan for deepfake misuse o

🛡️ Deepfake Watchdog 🤖

🏷️ Tags: mcp-server-track agent-demo-track

Video Link

A multimodal AI agent that detects deepfake content using face, voice, and video verification. Built with 🧠 DeepFace, 🗣️ SpeechBrain, 🎥 OpenCV, and accelerated with ☁️ Modal for serverless GPU-powered execution. Designed for the MCP HuggingFace's Hackathon.

🚀 Features

🧑‍🦱 Face Verification: Compare two faces and detect if they belong to the same person using DeepFace.
🎙️ Voice Verification: Determine if two audio samples are from the same speaker with SpeechBrain.
🎞️ Video Scan for Deepfakes: Scan a video (or YouTube link) and match all faces against a reference image.
📄 PDF Report: Generate a downloadable report with all scan results.
🧠 Runs as an MCP Agent: Compatible with Gradio MCP, fully pluggable into agentic workflows.
⚡ Modal Offloading: GPU-backed execution for DeepFace & SpeechBrain using Modal.

🧱 Architecture

Gradio UI ──> MCP Agent ──> Modal Functions └─> verify_faces_remote() └─> verify_voices_remote() └─> verify_faces_in_video_remote()

🧩 Tech Stack

Tool	Purpose
`Gradio`	Frontend UI + MCP integration
`Modal`	GPU offloading for ML workloads
`DeepFace`	Face verification
`SpeechBrain`	Voice verification
`OpenCV`	Video processing
`Pillow`	Image manipulation
`TensorFlow`	Required by DeepFace
`FFmpeg`	(Optional) for YouTube downloads

🛠️ Setup Instructions

1. Clone the repo

git clone https://huggingface.co/spaces/abetavarez/MCP-Hackathon-Deepfake-Watchdog
cd MCP-Hackathon-Deepfake-Watchdog

2. Create .env

cp .env.example .env
# Add your API keys or configs if required

3. Install dependencies (locally)

pip install -r requirements.txt

4. Run the app (with MCP)

modal run app.py

or for development::

python app.py

⚙️ Modal Setup

The app uses a single modal_app/backend.py file that defines:

verify_faces_remote

verify_voices_remote

verify_faces_in_video_remote

Each function is GPU-accelerated and runs independently on demand.

modal deploy modal_app.py

🧠 Agent MCP Integration

This app runs as an MCP server using:

demo.launch(mcp_server=True)

You can use this tool in your LlamaIndex, Hugging Face or LangChain agent as a remote service.

🏆 Hackathon Tracks ✅ Track 1: MCP Tool — Turned into a reusable tool with mcp_server=True ✅ Track 3: Agentic Demo — A fully capable AI agent for deepfake detection (LlamaIndex) ✅ Modal — Offloaded heavy tasks to Modal’s GPU cloud

🧱 Folder Structure

. ├── app.py # Main app entry point (Gradio + MCP) ├── llama_agent.py # LlamaIndex Agent implementation ├── tools.py # Tools for LlamaIndex Agent and MCP ├── modal_app/ │ ├── backend.py # Modal remote functions ├── detector/ │ ├── face.py # Local face verification logic │ ├── voice.py # Local voice verification logic │ └── video.py # Local video scan logic ├── reports/ │ └── pdf_report.py # PDF report generation ├── utils/ │ └── youtube_utils.py # YouTube download helper ├── README.md ├── requirements.txt └── .env.example

##🧑‍💻 Author Abraham Efren Tavarez GitHub - @AbeTavarez LinkedIn — @abrahametavarez

✨ Future Improvements

Add S3-compatible cloud storage or video ingest

Add Nebius S3 support for video uploads

Add RAG-based evidence summarizer

Enable auto-scan via background agent

Add QR code for mobile voice capture