Analysis of BERT Email Spam Classifier Against Adversarial Attacks

Email is still a preferred method of one-to-one communication among customers, businesses, and employees. As the usage of emails is increasing daily, the number of spam emails is also increasing. Many natural language models are developed to avoid the problem of spam emails. However, customers continue to report an increase in scams and attacks carried out through spam emails. Even though numerous spam filters have been suggested for identifying spam, they are weak and may be fooled by some expertly created adversarial samples. In this paper, we have taken the BERT classifier which has the best performance among other classifiers, and analyzed its performance when we inserted the adversarial samples. We have taken eight types of different adversarial attacks and analyzed the performance of the BERT Model. This analysis will be helpful in further studying defense methods against these attacks.