| --- |
| title: ImgAuth AI |
| emoji: 🛡️ |
| colorFrom: purple |
| colorTo: indigo |
| sdk: docker |
| app_port: 7860 |
| pinned: false |
| --- |
| |
| # ImgAuth AI — Image Authenticity Detector |
|
|
| 🛡️ **ImgAuth AI** is a state-of-the-art web application designed to detect AI-generated and manipulated images. Built with a "simple on the surface, powerful underneath" philosophy, it combines deep learning models with advanced digital forensics heuristics to deliver clear, binary verdicts: **Likely AI-Generated** or **Likely Authentic**. |
|
|
| Developed as a Major Project by **Team VisionGuard** (student team of 4). |
|
|
| --- |
|
|
| ## 🚀 Key Features |
|
|
| - **Binary Classification**: Simplified verdicts removing ambiguity ("Likely AI-Generated" or "Likely Authentic"). |
| - **Deep Learning Ensemble**: Combined predictions from 3 Hugging Face model pipelines: |
| - `umm-maybe/AI-image-detector` |
| - `dima806/ai_vs_real_image_detection` |
| - `Organika/sdxl-detector` |
| - **5 Forensic Heuristics**: Multi-layer analysis for technical validation: |
| 1. *Noise Kurtosis Analysis* (checks high-frequency noise distributions) |
| 2. *Deep Feature Inconsistency (DFI)* (checks patch-level consistency of Vision Transformer embeddings) |
| 3. *FFT Spectral Analysis* (identifies periodic artifacts in frequency domain) |
| 4. *Color Histogram Analysis* (detects synthetic pixel roughness/smoothness) |
| 5. *JPEG Ghost Analysis* (detects double compression artifacts in JPEG files) |
| - **AI Focus Areas (Explainability)**: Visual heatmaps showing ViT Attention Maps and Deep Feature Inconsistencies. |
| - **Collapsible Technical Drawer**: Advanced forensic signal logs, weights, and metrics available for researchers, while maintaining a clean, technical-jargon-free interface for everyday users. |
| - **Privacy First**: Fully stateless architecture; no images are stored permanently. Scanning history is saved only in local browser storage (`localStorage`). |
|
|
| --- |
|
|
| ## 👥 Meet Team VisionGuard |
|
|
| - **Vishal Chauhan** (Computer Science & Engineering, Project Lead) |
| - **Prince Mishra** (Computer Science & Engineering, Backend Developer) |
| - **Prince Dubey** (Computer Science & Engineering, Security & Testing) |
| - **Raksha** (Computer Science & Engineering, Frontend Developer) |
|
|
| --- |
|
|
| ## 🛠️ Technology Stack |
|
|
| - **Backend**: FastAPI, Uvicorn, PyTorch, Hugging Face Transformers, OpenCV, NumPy, SciPy |
| - **Frontend**: Vanilla HTML5, CSS3 (Modern dark-theme layout with purple gradients & glassmorphism), Vanilla JavaScript |
| - **Deployment**: Docker, Hugging Face Spaces |
|
|
| --- |
|
|
| ## 💻 Local Setup and Running |
|
|
| To run this application locally on your machine, follow these steps: |
|
|
| ### Prerequisites |
| - Python 3.10+ |
| - Pip package manager |
|
|
| ### Installation |
|
|
| 1. **Clone the repository**: |
| ```bash |
| git clone <repository-url> |
| cd imgauth-ai |
| ``` |
|
|
| 2. **Create and activate a virtual environment**: |
| - **Windows (PowerShell)**: |
| ```powershell |
| python -m venv .venv |
| .\.venv\Scripts\activate |
| ``` |
| - **macOS/Linux**: |
| ```bash |
| python -m venv .venv |
| source .venv/bin/activate |
| ``` |
| |
| 3. **Install dependencies**: |
| ```bash |
| pip install -r requirements.txt |
| ``` |
|
|
| 4. **Run the server**: |
| ```bash |
| python run.py |
| ``` |
| *The app will start running at:* `http://localhost:5000` |
|
|
| --- |
|
|
| ## 🐳 Running with Docker |
|
|
| Alternatively, build and run via Docker: |
|
|
| 1. **Build the image**: |
| ```bash |
| docker build -t imgauth-ai . |
| ``` |
|
|
| 2. **Run the container**: |
| ```bash |
| docker run -p 7860:7860 imgauth-ai |
| ``` |
| *Open browser to:* `http://localhost:7860` |
|
|
| --- |
|
|
| ## ⚖️ License & Attribution |
|
|
| - **Non-Commercial**: This project uses the `Organika/sdxl-detector` model, licensed under CC BY-NC 4.0. It is intended strictly for non-commercial educational and research purposes. |
| - **Model Attribution**: All deep learning classifications are handled by model weights published by the Hugging Face community. |
|
|
