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

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Machine Learning

Background:

• Email spam detection is a critical cybersecurity challenge due to continuous, nonstationary, and noisy data streams.
• Classifiers must adapt to evolving spam patterns without catastrophic forgetting.

Purpose of the Study:

• To investigate online email spam detection using hybrid Adaptive Resonance Theory (ART)-based neural models.
• To improve classifier robustness to noise and heterogeneous features in evolving data streams.

Main Methods:

• Utilized Fuzzy ARTMAP (FAM) for its stability-plasticity properties, augmented with preprocessing and postprocessing techniques.
• Compared concatenated and modular feature representations with FAM and Fusion ARTMAP (FusAM) architectures.
• Employed an ensemble majority voting method to reduce prediction errors.

Main Results:

• ART-based paradigms showed different accuracy-efficiency trade-offs.
• Feature extraction and ensemble methods consistently enhanced model robustness against noise.
• The ensemble FAM model achieved the best overall performance.

Conclusions:

• Hybrid ART-based models, particularly ensemble FAM, are effective for online email spam detection.
• Ensemble methods and feature engineering are crucial for robust performance in evolving, noisy data streams.
• Demonstrated the utility of ART models in adaptive cybersecurity applications.