Towards a reliable spam detection: an ensemble classification with rejection option

Despite significant investments in server infrastructure, spam continues to pose several challenges to the email ecosystem. These include ineffective detection and waste of computing resources. The constant need for periodic updates of spam detection model algorithms often contributes to these challenges. This paper proposes a reliable spam detection model that accounts for the dynamic nature of spam messages, which exhibit non-stationary behavior over time. The proposal is implemented in twofold. First, a shallow classifier selectively accepts only reliably classified spam messages to achieve a high detection rate. Second, unreliable classifications are forwarded to a deep-learning classifier. This ensures both reliability and high detection throughput. Experiments conducted on a novel dataset with 1.8 million real and valid spam messages collected over a decade demonstrate several key results: (i) the reliability of the model improves over time, (ii) autonomous identification of outdated models without human intervention, and (iii) the maintenance of high classification throughput.