TruthLens — Deepfake Detector

Detect AI-generated or manipulated images and videos instantly using a custom-trained Convolutional Neural Network.

---

Quick Start

# 1. Install dependencies
pip install -r requirements.txt

# 2. Run the app (model is already included)
python inference.py

# 3. Open browser
# http://127.0.0.1:5000

No training required — the pre-trained model (deepfake_detector_model.keras) is bundled.

---

What It Does

Upload any image or video and the model tells you in seconds whether it is REAL or FAKE (AI-generated / deepfake), along with a confidence score. For images, a Grad-CAM heatmap is overlaid to highlight which regions influenced the prediction.

---

How It Works

User uploads image/video
        ↓
Flask receives file → saves to /uploads
        ↓
Image: resized to 128×128 RGB → CNN inference → Grad-CAM overlay
Video: up to 16 frames sampled → per-frame CNN inference → mean score
        ↓
Sigmoid output:  ≥ 0.5 → FAKE   |   < 0.5 → REAL
        ↓
Result + confidence % returned to UI
        ↓
Uploaded file deleted

---

The Model

Property	Value
Architecture	EfficientNetB0 (ImageNet pretrained, fine-tuned)
Input size	224 × 224 × 3 (RGB)
Training	2-phase: frozen backbone → fine-tune top 30 layers
Output	Sigmoid (0 = Real, 1 = Fake)
Loss	Binary Crossentropy + label smoothing 0.1
Optimizer	Adam (Phase 1: 1e-3, Phase 2: 1e-5)
Expected Accuracy	~95%+
Training data	~140,000 images (70k Fake + 70k Real)
Framework	TensorFlow 2.x / Keras
Inference	Test-Time Augmentation (4-variant averaging)

---

Tech Stack

• Backend: Python, Flask
• ML: TensorFlow 2.x, Keras, OpenCV (Grad-CAM)
• Frontend: Vanilla HTML/CSS/JS (no framework, fully self-contained)

---

Project Structure

TruthLens/
├── inference.py                   # Flask web app + inference logic (entry point)
├── train.py                       # Training pipeline (model already trained)
├── deepfake_detector_model.keras  # Pre-trained model weights
├── requirements.txt               # Python dependencies
├── templates/
│   └── index.html                 # Frontend UI
├── static/
│   └── k.ico                      # Favicon
├── samples/                       # Demo test images
├── uploads/                       # Temp folder (auto-cleaned after each request)
└── data/                          # Dataset directory (gitignored)
    └── Dataset/
        ├── Train/
        ├── Test/
        └── Validation/

---

Supported Formats

Type	Extensions
Images	PNG, JPG, JPEG
Videos	MP4, MOV, AVI, MKV, WEBM

---

Training Your Own Model

# Place the dataset at data/Dataset/ with Train/, Test/, Validation/ subdirs
python train.py

This saves deepfake_detector_model.keras to the project root.

---

Credits

• Dataset: Trung-Nghia Le (Kaggle: deepfake-and-real-images)
• Model & App: TruthLens team