Fake accounts detection system based on bidirectional gated recurrent unit neural network

Abstract

Online social networks have become the most widely used medium to interact with friends and family, share news and important events or publish daily activities. However, this growing popularity has made social networks a target for suspicious exploitation such as the spreading of misleading or malicious information, making them less reliable and less trustworthy. In this paper, a fake account detection system based on the bidirectional gated recurrent unit (BiGRU) model is proposed. The focus has been on the content of users’ tweets to classify twitter user profile as legitimate or fake. Tweets are gathered in a single file and are transformed into a vector space using the GloVe word embedding technique in order to preserve the semantic and syntax context. Compared with the baseline models such as long short-term memory (LSTM) and convolutional neural networks (CNN), the results are promising and confirm that using GloVe with BiGRU classifier outperforms with 99.44% for accuracy and 99.25% for precision. To prove the efficiency of our approach the results obtained with GloVe were compared to Word2vec under the same conditions. Results confirm that GloVe with BiGRU classifier performs the best results for detection of fake Twitter accounts using only tweets content feature.