DEVELOPMENT OF TEXT-IMAGE SPAM DETECTION TECHNIQUES USING MULTI-MODAL RECURRENT NEURAL NETWORK AND CONVOLUTION NEURAL NETWORK TECHNIQUES

Abstract

ABSTRACT Spam emails are unsolicited message content shared through emails to several recipients using electronic devices. Despite the emergence of alternative forms of online communication which include social networking, sending and receiving emails has remained the most convenient and time efficient method of online communication. The increase in online transactions via email has led to a significant increase in the global number of email spam which has relatively become a critical problem in the area of computing. There have been numerous techniques of machine learning for identifying unsolicited email spam. Despite the significant improvements made in the number of existing literatures, there is no classification technique that has achieve 100% accuracy, each algorithm employs a limited number of features. Thus, determining the most appropriate technique is a critical task because their effectiveness needs to be weighed relative to their drawbacks. As a result, two deep learning techniques were explored and analyzed to identify both textual and image based email spam in this study. The study is aimed to analyze the effectiveness of Recurrent Neural Network (RNN) and Convolutional Neural Network (CNN) and develop a multi-modal architecture capable of detecting textual spam, image spam and mixed spam. Enron and Image Spam Hunter email datasets were used on the test size of 30% to obtain the performance of the model. The model was trained on text-image data and achieved an accuracy of 98% detection rate which indicates that the resultant model has outperforms the other models as compared to 85% achieved by Naïve Bayes, 95% achieved by Char-CNN and 97% achieved by Support Vector Machine (SVM) respectively.