Trinetra — Deepfake Forensic Analyzer

Trinetra is a comprehensive hybrid deepfake detection system that combines an optimized local AI model (EfficientNet-B4 + LSTM) with cloud-based verification (Reality Defender). It features a full forensic suite providing temporal anomaly detection, Explainable AI (XAI) using Grad-CAM, error level analysis, and geometric landmark jitter metrics.

Features

• Local Inference: Fast, on-device analysis using a hybrid EfficientNet-B4 + LSTM architecture for tracking spatial and temporal features.
• Cloud Verification: Integration with the Reality Defender API to serve as a robust fallback for verification analysis on demanding media.
• Deep Forensics Dashboard:
  • Executive Summaries
  • Grad-CAM Spatial Evidence for localized forgery highlighting
  • Temporal Anomaly and Attention Rollout Plots
  • Noise Residual Maps (Error Level Analysis)
  • Geometric Landmark Jitter Metrics
• Chrome Extension: Highlight regions of interest on any webpage and seamlessly scan for AI manipulation via a local background API.

Project Structure

• app.py: Gradio web interface for uploading and analyzing media interactively.
• api.py: FastAPI server acting as a bridge between the local Trinetra model and the Chrome extension interface.
• extension/: The unpacked Chrome extension containing content and background scripts.
• src/: Core Python modules responsible for inference, configurations, and forensic methodologies.
• model/: Directory to store model weights and checkpoints (e.g., best_model.pt).

Setup and Installation

1. Requirements

Ensure you have Python 3.9+ installed. Install all the required packages via pip:

pip install -r requirements.txt

2. Environment Variables

Create a .env file in the project's root directory. You will need to add your Reality Defender API key for the cloud verification features to function properly:

RD_API_KEY=your_api_key_here

3. Model Weights

The core AI model (best_model.pt) is excluded from this repository due to its large size (~250MB). To run the analysis features:

1. Obtain the model weights (e.g., from the project's releases or an external download link).
2. Place the best_model.pt file inside the model/ directory.
3. Ensure the directory structure looks exactly like this: Trinetra/model/best_model.pt.

4. Running the Interfaces

Gradio Web App:

To serve the interactive visual anomaly detection interface:

python app.py

Navigate to http://127.0.0.1:7860/ in your browser.

Local API (For Chrome Extension):

To serve the backend FastAPI endpoint to consume Chrome extension requests:

python api.py

Runs by default on http://127.0.0.1:8000/

WhatsApp Bot:

To run the deepfake scanner as a WhatsApp chatbot (requires Meta Cloud API credentials in your .env file):

cd backend
python whatsapp_bot.py

Runs on http://127.0.0.1:5000/ — expose this port via ngrok for the Meta webhook.

5. Chrome Extension Installation

1. Navigate to chrome://extensions/ in Chrome or any Chromium-based browser.
2. Enable "Developer Mode" in the top right corner.
3. Click "Load unpacked" and select the extension/ directory.

Models and Architecture

• Primary Pipeline: Extracts frame sequences or single images, applies Grad-CAM on CNN features (EfficientNet-B4), and calculates temporal divergence using an LSTM layer. Returns a localized heatmap, probability scoring, and extracted noise metrics.
• Cloud Layer: As a fallback and verification measure for unsupported scenarios, API calls to the Reality Defender endpoint confirm visual irregularities mapped against advanced cloud models.

License

This project is licensed under the MIT License. See the LICENSE file for details.