Adversarial Attacks against Machine Learning Classifiers: A Study of Sentiment Classification in Twitter

In recent years, the rapid evolution of social networks has played a significant role in people's lives. The wide use of social networks like Twitter, Facebook, and Instagram has shortened the communication distance between users. Besides communication, the social network has also become a platform for expressing one's thoughts, emotions, and opinions, becoming a medium for expressing attitudes towards entities such as products, services, events, etc. Therefore, the use of social media to attain various functions has become an investigation area in recent years. The extensive application of machine learning (ML) and natural language processing (NLP), and the improvements in computer performance, have been widely applied in various fields. In terms of social media, machine learning algorithms help in implementing functions like the sentimental analysis. To provide weighted sentiment scores to entities, the sentiment analysis system for text analysis integrates NLP and machine learning approaches. ML algorithms are sensitive to adversarial examples by adding perturbations or changes that are not detectable to the human naked eye and mislead the classifier to predict a false result. In this paper, we propose two attacks, LeetSpeakAttack and PhonologicalErrorAttack, that test the accuracy of a classifier. We also evaluate the classifiers' performance before and after the adversarial attack.