A novel defend against good word attacks

The good word attack is a common adversarial attack. The adversary defects spam filters by appending to spam some “good” words, which are words appearing frequently in legitimate emails but not in spam. The attacker expects add more “bad” words, which are the words that could distinctly convey the purpose of the advertisement to the emails. It can be perceived that the advertisement is more effective by adding more “bad” words since more information could be transmitted to the customers. As a result, forcing the attackers to diminish the number of “bad” words is an important research problem in good word attacks. In this paper, a novel method is proposed to force the attackers to diminish the number of “bad” words. Rather than only considering if a word contained in an email, the proposed method use the frequency of a word appeared in an email to simulate the adversary attack and the defense mechanism. Our proposed defense method is compared with different existing methods experimentally. The results show that our proposed have a better performance among those methods in term of accuracy.