Development of a Mobile Application for Deepfake Detection using the Xception Model

Abstract

Deepfake technologies are built with sophisticated AI and machine learning techniques such as Generative Adversarial Networks (GANs) that can generate highly realistic videos. Although deepfake technology is employed in various legitimate sectors like entertainment and education, they also pose a significant risk of being used for deceptive purposes, contributing to the spread of misinformation. To address this issue, a mobile application for detecting deepfake videos was developed. The study evaluated the performance of two deep learning models, Xception and EfficientNet-B7 on the DFDC dataset. The models were trained using Google Colab and were tested for accuracy in detecting deepfake content. The results showed that the Xception model outperformed the EfficientNet-B7 model, achieving an accuracy of 89.28% compared to 75.41% for EfficientNet-B7. Based on these findings, the Xception model was chosen to be implemented in the mobile application developed using MIT App Inventor and Flask API. The mobile app demonstrated 75% accuracy and a detection speed of 3.4891 seconds, representing a step forward in providing a robust and efficient deepfake detection solution for mobile platforms.