VerifAI

VerifAI is an Android application that detects AI-generated and deepfake images using a multimodal large language model backend. Users upload an image directly from their device; the app returns a verdict (real or fake) along with a natural-language explanation of the specific visual artifacts that indicate manipulation.

---

Overview

The proliferation of synthetic media has made it increasingly difficult to distinguish real photographs from AI-generated content. VerifAI addresses this problem by combining a production-grade Android client with a vision-language model inference backend, providing both a binary classification and an interpretable explanation for every image analysed.

The model powering the backend is FakeVLM (NeurIPS 2025), a multimodal large language model fine-tuned specifically for synthetic image detection across categories including human faces, animals, scenery, and documents.

---

Architecture (tentative)

User (Android App)
        │
        │  HTTPS  (image upload)
        ▼
  FastAPI Server  ──►  FakeVLM Model (HuggingFace Transformers)
        │
        │  JSON response (verdict + explanation)
        ▼
User (Android App)

Layer	Technology
Mobile client	Android (Kotlin)
Authentication	Firebase Auth
API server	FastAPI (Python)
ML inference	FakeVLM via HuggingFace Transformers
Model	lingcco/fakeVLM (LLaVA-based, 7B parameters)

---

Features

• Upload any image from your Android device for AI/deepfake detection
• Receive a plain-language explanation of detected visual artifacts
• Firebase authentication for secure user access
• REST API backend compatible with future GCP deployment
• 4-bit quantized inference for consumer GPU compatibility

---

Repository Structure

team-tj/
├── android/                    # Android application (Kotlin)
│   └── app/
│       └── src/main/           # Kotlin source, layouts, resources
├── backend/                    # FastAPI inference server
│   ├── server.py               # Entry point (python server.py)
│   └── app/                    # API routes, model handler, config
└── README.md

---

Backend Setup

Prerequisites

• Python 3.10
• NVIDIA GPU (12 GB VRAM minimum; 16 GB recommended for production)
• CUDA 12.x
• WSL2 (if running on Windows) or native Linux

Installation

# Clone the repository
git clone https://github.com/hcp-uw/team-tj.git
cd team-tj

# Create and activate the conda environment
conda create -n fakevlm python=3.10 -y
conda activate fakevlm

# Install CUDA toolkit (required for compilation)
conda install -c nvidia cuda-toolkit cuda-runtime cuda-nvcc -y

# Install Python dependencies
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
pip install transformers accelerate bitsandbytes
pip install fastapi uvicorn python-multipart Pillow

# Download the FakeVLM model weights
cd backend
huggingface-cli download lingcco/fakeVLM --local-dir ./FakeVLM/fakeVLM_model

Running the Server

conda activate fakevlm
cd backend
python server.py

The server starts on http://0.0.0.0:8080.

API Reference

POST /analyze

Accepts a multipart image upload and returns a detection result.

Request

curl -X POST http://localhost:8080/analyze \
  -F "file=@image.jpg"

Response

{
  "status": "success",
  "result": "This is a fake image. The image exhibits underlying characteristic inconsistencies in its features that suggest it is artificially created."
}

GET /health

Returns server status.

{ "status": "ok" }

---

Android Setup

Prerequisites

• Android Studio (latest stable)
• Android SDK 26+
• A Firebase project with Authentication enabled

Configuration

1. Create a Firebase project at console.firebase.google.com
2. Enable Email/Password authentication
3. Enable Cloud Storage and Firestore Database
4. Deploy security rules from firebase/storage.rules and firebase/firestore.rules (Firebase Console → Rules)
5. Download google-services.json and place it in android/app/
6. Set your backend URL in android/local.properties (optional; defaults to emulator host):

   api.base.url=http://10.0.2.2:8080

   For a physical device, use your computer's LAN IP, e.g. http://192.168.1.10:8080

Firebase PERMISSION_DENIED

If the app shows PERMISSION_DENIED: Missing or insufficient permissions, the signed-in user is fine but Firestore or Storage is rejecting writes. Typical causes:

1. Rules not published — In Firebase Console, open Firestore → Rules and Storage → Rules, then paste the contents of firebase/firestore.rules and firebase/storage.rules from this repo and Publish.
2. Firestore or Storage not enabled — Create a default database and default bucket if the console prompts you.

Until rules allow users/{uid}/…, cloud history may stay empty; the app still runs inference if the Python backend is up.

Building

Open the project in Android Studio and run on a device or emulator (API 26+).

---

Model

VerifAI uses FakeVLM, a multimodal large language model presented at NeurIPS 2025 by OpenDataLab. It is built on the LLaVA architecture and fine-tuned on the FakeClue dataset, which contains fine-grained natural-language annotations of visual artifacts across multiple image categories.

• Paper: FakeVLM — NeurIPS 2025
• Model weights: lingcco/fakeVLM on HuggingFace
• Evaluated on FakeClue, LOKI, and DD-VQA benchmarks

---

Roadmap

• GCP deployment with vLLM for production throughput
• Priority request queueing for concurrent users
• Confidence score in API response
• In-app image history and result storage
• Support for video frame analysis

---

Acknowledgements

• FakeVLM by OpenDataLab
• LLaVA — base architecture
• vLLM — planned production inference engine

---

License

This project is for academic and research purposes. See LICENSE for details.