ImageFake: An Ensemble Convolution Models Driven Approach for Image Based Fake News Detection

Abstract

Social media usage has shot up over the past decade which gives us many wonderful opportunities to showcase ourselves, our skills, and field of competences. In short, the facial nature of social media raised opportunities to create and promote one’s own brand to gain influence in the market. Besides exploiting these social media content uplifted the negative impacts of spreading false information as well. As a consequence, falsehood detection in terms of information gain diverted researchers’ attention and the same issue has been addressed in this work. We present an ImageFake ensemble model which makes use of abundant pre-trained CNN models to discover the usage of multimedia features for image-based fake news detection and classification. The pre-trained models used for fake news detection and classification are VGG-16, VGG-19, Inception v3, SqueezeNet, and ResNet-101, and finally the bagging ensemble model is used for selecting the best of the bunch. MediaEval 2015, a Twitter dataset is used for the experiment and to validate multimedia feature usage. The performance of pre-trained CNN models on the particular visual domain is performed and compared and out of all best-considered models in terms of accuracy and execution time is ResNet-101 which is able to achieve 96% training accuracy. The ensemble model is able to get training and validation accuracy 97% and 66% respectively.