Fighting Evolving Spam With ARTMAP Models: A Noise-Resilient Online Detection Framework

Email spam detection is a core cybersecurity challenge in which data samples (emails) normally arrive as a continuous, nonstationary stream. The corresponding data samples are often noisy or deliberately obfuscated. To effectively classify a variety of emails in such settings, a practical classifier must adapt online to changing spam patterns without catastrophically forgetting previously learned patterns. This study investigates online email spam detection and classification using hybrid Adaptive Resonance Theory (ART)-based neural models. Specifically, Fuzzy ARTMAP (FAM) is adopted as the underlying classifier due to its stability-plasticity properties. To improve robustness to noise and heterogeneous features, we augment FAM with preprocessing and postprocessing techniques, and systematically compare both concatenated and modular feature representations with FAM and Fusion ARTMAP (FusAM) architectures. We further employ an ensemble majority voting method to reduce prediction errors under noisy, evolving data streams. Comprehensive experiments on email spam data samples with and without feature/attribute noise indicate that: 1)the concatenated and modular ART-based paradigms exhibit different accuracy-efficiency trade-offs, and 2)feature extraction and ensemble methods consistently improve model robustness. The best overall performance is achieved by an ensemble FAM model, demonstrating its usefulness for evolving email spam detection in cybersecurity applications.