A Multimodal Spam Filtering System for Multimedia Messaging Service

Abstract

Filtering systems that detect spam by a single modality have increased over the past few years for Short Message Services (SMS). With the growth of marketing via Multimedia Messaging Services (MMS), a spammer can evade detection by injecting multimodal junk information into MMS to decrease systems recognition based on a single modality. Due to this situation, we propose in this paper, a powerful spam MMS filtering system, with a multi-modality fusion technique, it can detect spam whether it is hidden in text or images. The architecture of the proposed system combines along Short-Term Memory (LSTM) model and a Convolutional Neural Network (CNN) model to filter spam MMS. Based on the text and image portions of an MMS, the LSTM and the CNN calculate two classification probabilities. Afterwards, to determine whether the MMS is spam or not, these values are incorporated into a fusion model. Based on the results of the experiments, the overall accuracy is 98.56% for the MMS dataset, 96.97% for the text-based dataset, and 97.98% for the image-based dataset. We have compared the proposed system with some relevant ones and we have found that it performs better in many criteria such as precision, f1-score, recall, and accuracy and gives an improvement between 1 and 5% in smishing detection.