Abstract

This paper proposes to use the Naïve Bayes-based algorithm for phishing detection, specifically in spam emails. The paper compares probability-based and frequency-based approaches and investigates the impact of imbalanced datasets and the use of stemming as a natural language processing (NLP) technique. Results show that both algorithms perform similarly in spam detection, with the choice between them depending on factors such as efficiency and scalability. Accuracy is influenced by the dataset configuration and stemming. Imbalanced datasets lead to higher accuracy in detecting emails in the majority class, while they struggle to classify minority-class emails. In contrast, balanced datasets yield overall high accuracy for both spam and ham email identification. This study reveals that stemming has a minor impact on algorithm performance, occasionally decreasing in accuracy due to word grouping. Balancing the dataset is crucial for improving algorithm performance and achieving accurate spam email detection. Hence, both probability-based and frequency-based Naïve Bayes algorithms are effective for phishing detection using balanced datasets. The frequency-based approach, with a balanced dataset and stemming, achieves a balanced performance between recall and precision, while the probability-based method with a balanced dataset and no stemming prioritises overall accuracy.