On the evaluation of email spam filters in adversary environment

Abstract

This paper proposes a novel framework for assessing the robustness of email spam filters against adversarial attacks. We introduce a methodology to quantify vulnerability by simulating attacker strategies that deliberately modify the training and/or test sets to maximize performance degradation. This degradation is measured using the Area Under the Curve (AUC) within the high-sensitivity region (0.9-1.0) of the Receiver Operating Characteristic (ROC) curve. Our experimental results demonstrate that Random Forest-based filters exhibit superior robustness compared to existing approaches, representing a significant advancement in adversarial-resistant spam filtering. The proposed evaluation framework provides critical insights into filter vulnerabilities under optimal attack conditions, addressing a key gap in current spam filter assessment methodologies.