Fake Image Detection on Social Networks using Hybrid Deep Learning Methodologies

Abstract

The availability of a large amount of accessible content on social media, along with sophisticated tools and processing infrastructure at a low cost, has made it extremely easy for users to produce sophisticated forgeries that have the potential to propagate hoaxes and misinformation. The ease with which anyone can use these tools to produce propaganda could lead to fear and anarchy because of the rapid growth of these technologies. Therefore, it has now become essential to have a mechanism to identify genuine and fake images. In this regard, this research aims to present a new hybrid classification model based on the Restricted Boltzmann Machine (RBM) and Recurrent Neural Network (RNN), which utilizes Deep Learning (DL) techniques to present an automated solution for detecting fake images on social media platforms. Firstly, the input images are pre-processed through image resizing and synthesizing its Error Level Analysis (ELA). Then, ResNet-152, Alex-Net, and Squeeze-Net are utilized to extract the features of the image. Moreover, a hybrid metaheuristic algorithm called Salp Inspired Reptile Search Optimization (SIRSO) is presented to optimize the hyperparameters of the suggested model. The suggested model achieves an Accuracy of 98.99%, and the model is implemented in Python programming language. From the results, it is clear that the suggested model is performing well when compared to SOTA methods.