Abstract
Twitter is one of the most popular online social networks for spreading propaganda and words in the Arab region. Spammers are now creating rogue accounts to distribute adult content through Arabic tweets that Arabic norms and cultures prohibit. Arab governments are facing a huge challenge in the detection of these accounts. Researchers have extensively studied English spam on online social networks, while to date, social network spam in other languages has been completely ignored. In our previous study, we estimated that rogue and spam content accounted for approximately three quarters of all content with Arabic trending hashtags in Saudi Arabia. This alarming rate, supported by autonomous concurrent estimates, highlights the urgent need to develop adaptive spam detection methods. In this work, we collected a pure data set from spam accounts producing Arabic tweets. We applied lightweight feature engineering based on rogue content and user profiles. The 47 generated features were analyzed, and the best features were selected. Our performance results show that the random forest classification algorithm with 16 features performs best, with accuracy rates greater than 90%.