TransDFD: A Deepfake Detection System of Mesoscopic level Deepfake-guard-AI

This research introduces TransDFD, a novel deepfake detection system developed at the mesoscopic level using a deep learning approach and deployed as an application named "Deepfake-guard-AI" for the public. The model enhances the original MesoNet-4 architecture by integrating transfer learning with some additional layers, optimizing the performance, and also improving the system for an Indian dataset. The added layers, such as Conv2D, Batch Normalization, MaxPooling2D, and LeakyReLU, improve adaptability and precision. The system successfully identifies the videos that were altered using methods such as Deepfake and Face2Face manipulations. TransDFD leverages pre-trained MesoNet weights and fine-tunes t hem to detect deepfakes by capturing precise pixel-level artifacts. The proposed model analyzes video frames, classifying them as real or fake based on the learned features. A Web application built using Streamlit enables users to upload videos for real-time deepfake detection, providing classification results through a user-friendly interface. Our approach offers a scalable solution that achieves high accuracy while using low computational resources, making it applicable for both real-time detection and post-event analysis in security and media integrity contexts. This research demonstrates a significant advancement in deepfake detection, offering a robust and adaptable model capable of addressing current challenges in the identification of manipulated video content and efficient detection of such videos.