AI Content Forensics Suite (ACFS)

A research-prototype application for detecting AI-generated content in both text and images with explainability and reproducible forensic reporting.

Features

• Multi-modal Detection: Detects AI-generated images (ViT/DeiT) and text (RoBERTa/DeBERTa)
• Forensic Analysis: ELA, EXIF metadata (images), stylometry, perplexity (text)
• Explainability: Grad-CAM/Attention Rollout (images), Integrated Gradients/SHAP (text)
• Fusion Layer: Calibrated probability averaging for robust classification
• Forensic Reports: PDF reports with hashes for reproducibility
• Deployment: FastAPI backend + Streamlit frontend, Dockerized

Project Structure

ACFS/
├── backend/                 # FastAPI backend
│   ├── api/                # API endpoints
│   ├── models/             # Model inference modules
│   ├── forensics/          # Forensic analysis tools
│   ├── explainability/     # Explainability modules
│   ├── fusion/             # Fusion layer
│   └── reporting/          # PDF report generation
├── frontend/               # Streamlit UI
├── training/               # Model training scripts
├── data/                   # Dataset directory (DVC managed)
├── models/                 # Saved model checkpoints
├── reports/                # Generated forensic reports
├── tests/                  # Unit and integration tests
├── docker/                 # Docker configuration
└── config/                 # Configuration files

Installation

Local Setup

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Initialize DVC
dvc init

Docker Setup

# Build Docker image
docker build -f docker/Dockerfile -t acfs:latest .

# Run container
docker run -p 8000:8000 -p 8501:8501 acfs:latest

Usage

Start Backend (FastAPI)

cd backend
uvicorn main:app --host 0.0.0.0 --port 8000 --reload

Start Frontend (Streamlit)

streamlit run frontend/app.py --server.port 8501

Train Models

# Train image detection models
python training/train_image_models.py --model vit --epochs 10

# Train text detection models
python training/train_text_models.py --model roberta --epochs 5

API Endpoints

• POST /api/analyze/image - Analyze image for AI generation
• POST /api/analyze/text - Analyze text for AI generation
• GET /api/report/{report_id} - Download forensic report
• GET /api/models/status - Check model status

Reproducibility

All models, datasets, and reports include:

• SHA256 hashes for integrity verification
• Timestamps for audit trails
• DVC versioning for datasets
• Dockerized environment for consistency

Tech Stack

• ML/DL: PyTorch, HuggingFace Transformers, timm
• Explainability: Captum, SHAP, pytorch-grad-cam
• Forensics: OpenCV, ExifRead, Pillow
• Backend: FastAPI, Pydantic
• Frontend: Streamlit
• Reporting: ReportLab
• Reproducibility: Docker, DVC

License

MIT License

Authors

Research Prototype - AI Content Forensics Suite Team