IMPROVED FAKE IMAGE DETECTION AND CLASSIFICATION USING XCEPTION MODEL

Abstract

Recently, the creation of hyper-realistic fake faces using deep learning and machine learning methods has risen, resulting in a higher occurrence of fake images. These synthetic images, capable of convincingly imitating real human faces, present significant threats in various areas, including security and privacy, as well as media credibility and digital trust. This paper proposes a custom Xception deep learning model for fake image detection and classification. The base Xception model, pretrained on ImageNet, is included without its fully connected (i.e., top) layers to allow the model to utilize learned features with a focus on binary classification for our specific deepfake and real image dataset. The input layer accepts images of 128x128 pixels with three color channels providing compact input for reduced computational load. The proposed framework achieved the highest accuracy of 97.85% validation accuracy over training accuracy of 98.61% for both the deepfake and real image datasets. It also achieves high performance on various evaluation metrics with a Precision score of 0.95, a recall score of 0.79 and F1-Score of 0.86. The experimental results obtained demonstrated that the suggested method surpassed other leading fake image discriminators regarding performance, and it can assist cybersecurity experts in combating cybercrimes related to deepfakes.