🛡️ DeepGuard AI — Deepfake Detection System

A real-time deepfake detection system built with EfficientNetV2B0 + CBAM Attention, deployed as both a Chrome Extension and a Streamlit web app for AI-generated content detection — directly in the browser, with zero data leaving your device.

Final Model Performance:

Metric	Score
Accuracy	88.28%
AUC-ROC	0.9571
Average Precision	0.9585
F1 Score (optimized threshold)	0.8880

Evaluated on a balanced test set of 20,000 samples (10,000 real, 10,000 fake), trained on ~140K real/fake face images.

---

🌐 Chrome Extension

Features

• Real-time page scanning — Automatically detects AI-generated images and videos on any web page
• 100% local inference — All analysis runs on-device using TensorFlow.js; no data ever leaves your machine
• Smart detection — MutationObserver-based scanning catches dynamically loaded content (infinite scroll, SPAs)
• Video frame analysis — Extracts and analyzes video frames at configurable intervals
• Premium UI — Glassmorphism popup dashboard with real-time stats, detection history, and configurable settings
• In-page overlays — Non-intrusive badges on scanned images showing Real/AI-Generated status with confidence scores
• User-configurable — Toggle auto-scan, adjust sensitivity, show/hide overlays, enable/disable video scanning

Extension Setup

1. Convert the trained model to TensorFlow.js format:

   pip install tensorflowjs
   python scripts/convert_model.py

   This produces extension/model/model.json + weight .bin files.

2. Load the extension in Chrome:

   • Open chrome://extensions/
   • Enable Developer mode (top-right toggle)
   • Click Load unpacked
   • Select the extension/ folder

3. Usage:

   • Click the 🛡️ DeepGuard AI icon in the toolbar
   • Click "Scan This Page" to analyze all images/videos
   • Toggle Auto-Scan in settings to scan pages automatically
   • Hover over badges on images to see detailed confidence scores

Extension Architecture

extension/
├── manifest.json          # Manifest V3 configuration
├── background.js          # Service worker (message routing, state management)
├── content/
│   ├── content.js         # DOM scanner, MutationObserver, overlay injection
│   └── overlay.css        # In-page glassmorphism badges
├── offscreen/
│   ├── offscreen.html     # Hidden document for TF.js inference
│   └── offscreen.js       # Model loading, preprocessing, prediction
├── popup/
│   ├── popup.html         # Dashboard UI
│   ├── popup.css          # Premium glassmorphism styles
│   └── popup.js           # Settings, stats, scan control
├── lib/
│   └── tf.min.js          # TensorFlow.js library (~1.4 MB)
├── model/
│   ├── model.json         # Converted model topology
│   └── *.bin              # Weight shards (~10-15 MB)
└── assets/
    └── icons/             # Extension icons (16, 48, 128px)

---

💻 Streamlit Web App

For deeper, manual analysis beyond the extension's automatic scanning. Powered by the same detection engine.

Modes

• Image Analysis — Upload a single image and get an instant Real/AI-Generated verdict with confidence score
• Video Analysis — Upload a video for full per-frame breakdown, aggregated prediction, and fake-frame percentage
• Live Capture — Use your webcam to capture a photo and get instant on-the-spot deepfake verification

Each mode returns face detection results, confidence scoring, and a downloadable result summary/card.

Launch

streamlit run app.py

---

🔬 ML Training Pipeline (Python)

Folder Structure

• data/raw: Original dataset images/videos (managed by download_data.py)
• data/processed: Cropped faces and preprocessed data
• models/: Saved model weights (.h5, .keras)
• src/: Source code for preprocessing, training, and inference
• notebooks/: Jupyter notebooks for experimentation
• scripts/: Model conversion utilities
• download_data.py: Script to download and set up the dataset

Environment Setup

1. Create a virtual environment:

   python -m venv venv

2. Activate the environment:
   • Windows: .\venv\Scripts\activate
   • Mac/Linux: source venv/bin/activate
3. Install dependencies:

   pip install -r requirements.txt

Dataset Setup

1. Kaggle API:
   • Go to Kaggle Account
   • Click "Create New API Token" to download kaggle.json
   • Place kaggle.json in C:\Users\<YourUser>\.kaggle\ OR in this project folder
2. Download Data:

   python download_data.py

   This downloads the dataset and organizes it into data/raw/real and data/raw/fake.

Data Preprocessing

Extract faces from videos/images using MTCNN:

python src/preprocess.py

This populates data/processed/real and data/processed/fake.

Model Training

Train the EfficientNetV2B0 + CBAM model with two-phase training (frozen backbone → fine-tuning):

python src/train.py

This saves the model to models/deepfake_detector.keras.

Evaluation

Generate a comprehensive evaluation report (confusion matrix, ROC curve, PR curve, threshold analysis):

python src/evaluate.py

---

🧠 Model Architecture

• Base Model: EfficientNetV2B0 (pre-trained on ImageNet)
• Preprocessing: MTCNN for face detection and extraction
• Attention: CBAM (Convolutional Block Attention Module) — focuses on facial regions like eyes, skin texture, and lighting seams
• Custom Head: GAP → Dense(512) → BN → Dropout → Dense(256) → BN → Dropout → Sigmoid
• Training Strategy: Two-phase (Phase 1: frozen backbone, classification head only → Phase 2: fine-tuning top layers)
• Loss Function: Binary Focal Crossentropy
• Metrics Tracked: Accuracy, AUC-ROC, Precision, Recall, F1-Score
• Deployment Optimization: Float16 quantization for TensorFlow.js, enabling fast browser-based inference via WebGL

---

📊 Results Summary

Metric	Value
Test Set Size	20,000 samples (balanced)
Accuracy	88.28%
AUC-ROC	0.9571
Average Precision	0.9585
F1 @ 0.5 threshold	0.8782
Best F1 (optimized threshold = 0.438)	0.8880
True Positives / True Negatives	8,450 / 9,206
False Positives / False Negatives	794 / 1,550

---

📄 Research Paper

This project has been submitted to ICIAEM-2026 for academic review.

---

Requirements

• Python 3.10+
• TensorFlow 2.12+
• Chrome 120+ (for extension)
• See requirements.txt for full dependency list

---

🔮 Future Scope

• Temporal (RNN/LSTM) modeling for video-native deepfake detection
• Larger, more diverse training datasets for improved generalization
• Cross-browser and mobile support
• Explainable AI features to visualize why content is flagged as fake

---

Author

Built by Prathamesh Zagade (@prathamxz) as a final year B.E. project (Artificial Intelligence & Data Science).