Detecting Spam Images with Embedded Arabic Text in Twitter

Abstract

Detecting image-based spam in Online Social Networks (OSNs), such as Facebook and Twitter, is an ongoing problem. Spam is prevalent in all forms of online communication (such as email and the web) However, researchers' and practitioners' attention has increasingly shifted to spam in OSNs, due to the growing number of spammers and the possible negative effects on users. There are different types of spam messages that can be found in OSNs. Spam images, which are images embedded with malicious text, are one of the most difficult types of spam to tackle. Processing images overwhelms classifiers and affects detection performance. Consequently, spammers take advantage of this issue to launch more sophisticated attacks, such as evasion attacks. After observing some Arabic trending hashtags and topics in Twitter, a substantial amount of image-based spam was found. Thus, this paper proposes an approach for detecting image-based spam with Arabic text in Twitter through using Deep Learning (DL) techniques. In this paper, an Efficient and Accurate Scene Text Detector (EAST) and Convolutional Recurrent Neural Network (CRNN) models were used for text detection and text recognition. After the text extraction step, a blacklist and whitelist approach was applied for classifying text as either spam or non-spam. The proposed text classification technique is adaptable and robust against some text classification attacks.