AI Image & Video Detector

Detects whether an image or video is AI-generated using:

• EfficientNet-B4 neural classifier
• FFT frequency analysis for artifact detection
• Ensemble scoring (80% neural + 20% FFT)

---

Quick Start

1. Install dependencies

pip install -r requirements.txt

2. Run the app

python app.py

Open http://localhost:5000 in your browser.

---

Project Structure

ai-detector/
├── app.py              # Flask web server
├── model.py            # Model loading + inference
├── train.py            # Fine-tuning script
├── requirements.txt
├── models/
│   └── detector.pth    # (optional) fine-tuned weights
├── templates/
│   └── index.html      # Frontend UI
└── data/               # (for training only)
    ├── train/
    │   ├── real/
    │   └── fake/
    └── val/
        ├── real/
        └── fake/

---

Training on your own dataset

1. Organize images into data/train/real, data/train/fake, data/val/real, data/val/fake

2. Run:

python train.py

The best checkpoint is saved to models/detector.pth and auto-loaded by the app.

Recommended public datasets

Dataset	Link
CIFAKE	https://www.kaggle.com/datasets/birdy654/cifake-real-and-ai-generated-synthetic-images
CNNDetection	https://github.com/peterwang512/CNNDetection
GenImage	https://github.com/GenImage-Dataset/GenImage
FaceForensics++	https://github.com/ondyari/FaceForensics (for deepfake video)

---

Upgrading the model

Swap efficientnet_b4 in model.py for any timm model, or load weights from:

• UniversalFakeDetect: strongest for diffusion images
• FaceForensics++: best for deepfake faces in video

---

Notes

• No fine-tuned weights = ImageNet pretrained only (weaker, but still runs)

• Video analysis samples 1 frame every 15 frames (configurable in model.py)

• Heavy JPEG compression degrades detection accuracy

• Confidence shown is for AI Generated — lower = more likely real

• The Working Link (" https://maheshepili.github.io/Ai-Detector/ ")