A Comparative Analysis of Deep Neural Network Models for Deepfake Image Detection Generated using AI

Deepfake consists of two words - “deep” and “fake,” where deep refers to the deep neural network and fake indicates the manipulation of data. The gradual increase in usage of AI-generated images poses serious concerns and threats as these kinds of images challenge the authenticity of images and videos. This manipulation of pictures should be addressed to reduce their misuse in spreading misinformation, identity theft, and other malicious activities. This study mainly focuses on developing a model that can help detect the use of deepfake images. We used convolutional neural networks (CNNs) together with advanced architectures ResNet50 and XceptionNet to classify whether an image is real or fake. We trained our model on two different datasets and compared their accuracies. These datasets contained authentic as well as deepfake images which are used for the model's training and testing. Comparing the three models, it is found that the XceptionNet architecture performed classification with an accuracy of 84%. To enhance detection accuracy, preprocessing techniques are also applied to extract meaningful features and reduce noise in the data.