Reproduce “Phishing Detection through Email Embeddings” on OpenResearch

About this paper

Luis Felipe Gutiérrez, Faranak Abri, Miriam Armstrong, Akbar Siami Namin, Keith S. Jones

An investigation into phishing detection demonstrates the efficacy of Doc2Vec email embeddings for distinguishing malicious emails, even when their content closely mimics legitimate communications. The approach, particularly when combined with linear dimensionality reduction, enabled machine learning models to achieve up to 91.6% accuracy.

Key insights

• Email embedding techniques like Doc2Vec enable machine learning models to be "content-aware," effectively distinguishing phishing emails from legitimate ones despite intentionally similar content cues.
• Linear Principal Component Analysis (PCA) significantly improved classification performance when reducing Doc2Vec embeddings to 2 dimensions, outperforming both the full 20-dimensional space and non-linear RBF Kernel PCA.
• The study indicated that the most discriminative information for classification resided within the first two linear PCA components, which captured approximately 25% of the total variance, yet yielded the highest classification accuracy.

Summary by alphaXiv.