lakshman-a01/deepfake-detection-system-deeplearning

Purpose and Runtime Entry Point

This is a Python deepfake-detection application for uploaded images and videos. It classifies detected face crops as real or fake using a trained deep-learning workflow developed in notebooks. The intended runtime entry point is deep/app.py; related launch/prototype modules are deep/main.py and deep/new.py.

Web Application and User Flow

The application is a server-rendered web interface: users navigate from home/login/registration views to image or video upload forms, then receive corresponding display/result pages. Template boundaries are in deep/templates/home.html, deep/templates/Upload_image.html, deep/templates/Upload_Video.html, deep/templates/Display_image.html, and deep/templates/Display_Video.html. Styling and bundled media are served from deep/static/style.css and deep/static/.

Upload and Media Storage Boundary

Uploaded media is persisted under deep/uploads/, separating user-provided content from static demonstration assets. Video examples and uploads present in the repository indicate the application accepts MP4 inputs, while image handling is supplied through the image upload/display templates. Relevant paths: deep/uploads/tomc_orginial.mp4, deep/uploads/tomo.mp4, and deep/static/Very_realistic_Tom_Cruise_Deepfake___AI_Tom_Cruise.mp4.

Video Detection Pipeline

For video inputs, the core processing flow is: ingest video → extract frames → locate faces → crop face regions → classify each crop with the deep-learning model → present an aggregate/display result. Face localization uses the bundled OpenCV Haar cascade at deep/haarcascade_frontalface_default.xml. The model experimentation and training pipeline are represented by deep/model/Preprocessing.ipynb, deep/model/deepfake_detection_train.ipynb, and deep/model/Prediction.ipynb.

Image Detection Pipeline

For static images, the pipeline bypasses frame extraction: upload image → detect/crop a face → invoke the classifier → render the image result. This is exposed separately from video processing through deep/templates/Upload_image.html and deep/templates/Display_image.html, allowing modality-specific request handling and presentation.

Model and ML Boundary

Model preprocessing, training, and prediction are notebook-driven rather than packaged as a dedicated Python model module. deep/model/Preprocessing.ipynb prepares facial inputs, deep/model/deepfake_detection_train.ipynb trains the detector, and deep/model/Prediction.ipynb performs inference-oriented work. Application code in deep/app.py is therefore the integration boundary between web requests, OpenCV face detection, filesystem media, and the trained-model inference logic.

Dependencies and Deployment Assumptions

The project runs in a local Python/Conda environment with dependencies declared in deep/requirements.txt. It depends materially on deep-learning libraries and computer-vision support for classification and face detection, plus a Python web-server framework for browser interaction. No cloud infrastructure, remote model-serving endpoint, database, or external API is indicated; media and inference appear to be local.