Towards a Reliable Spam Detection: An Ensemble classification with rejection option

Abstract

Abstract Many issues are faced in the email environment due to Spam, such as bottlenecks in the email gateways despite substantial investments in servers' infrastructure, wasted computational resources, and ineffective detection despite the demand for frequent spam model updates. This paper proposes a reliable detection model to deal with the non-stationary behavior of spam messages over time. A high detection rate is provided in a shallow classifier wherein only reliable spam message classification is accepted. Unreliable classifications are rejected and forwarded to a deep learning classifier, providing reliability and a high detection throughput. Experiments performed on a new dataset with 1,898,843 real and valid spam messages stored for over ten years show that they (i) can improve its reliability over time, (ii) detect outdated models without human assistance, and (iii) provide a high classification throughput rate.