Enhancing Email Filtering Systems: Mitigating Adversarial Attacks with Advanced Defense Mechanisms

Abstract

Email filtering systems, relying on Machine Learning (ML) and Natural Language Processing (NLP), are increasingly vulnerable to adversarial attacks, where malicious actors subtly alter email content to bypass detection. These attacks exploit weaknesses in current models, making it difficult for conventional filters to identify phishing or spam threats. Existing methods often fail to detect such adversarial examples, posing significant security risks. This paper proposes a robust email classification system capable of accurately distinguishing spam from legitimate (ham) emails, even under adversarial conditions. The system evaluates multiple traditional machines learning algorithms, including Random Forest, Support Vector Machine, Logistic Regression, and Naive Bayes, and compares their performance before and after adversarial attacks. Six types of text-based adversarial attacks are applied: character swap, homoglyph replacement, synonym substitution, word-level noise, backtranslation, and paraphrasing. Adversarial training considerably increases classification robustness, according to evaluation on the Enron email dataset. The suggested ensemble method mitigates the performance loss commonly seen in adversarial circumstances while maintaining excellent accuracy.