AI-based phishing detection in email

Abstract

Phishing emails are one of the biggest web threats that deceive individuals into providing confidential information by claiming to belong to a trusted party. Such traditional security measures as rule-based filters or a blacklist are often incapable of keeping pace with these quickly-evolving attacks. In this thesis, the role of Artificial Intelligence (AI) is taken into account to provide more intelligent and adaptive protection. Two deep learning models have been compared utilizing one specially designed dataset of real phishing and authentic email: Long Short-Term Memory (LSTM) networks and Bidirectional Encoder Representations of Transformer (BERT). The LSTM model demonstrated strong performance, with an average accuracy of 98.75% and within the margin of perfection, as compared to BERT, which showed a lot of promise and required additional data and fine-tuning. These results suggest that AI can not only detect more elaborate patterns of phishing attempts itself, but also be more effective in changing with the circumstances than traditional methods. At the same time, the study highlights the existing problems of privacy, adversarial attacks and how there exists a need for transparent and trusting AI systems.