Abstract
Phishing is an identity theft that employs social engineering methods to get confidential data from unwary users. A phisher frequently attempts to trick the victim into clicking a URL that leads to a malicious website. Many phishing attack victims lose their credentials and digital assets daily. This study demonstrates how the performance of traditional machine learning (ML)-based phishing detection models deteriorates over time. This failure is due to drastic changes in feature distributions caused by new phishing techniques and technological evolution over time. This paper explores continual learning (CL) techniques for sustained phishing detection performance over time. To demonstrate this behavior, we collect phishing and benign samples for three consecutive years from 2018 to 2020 and divide them into six datasets to evaluate traditional ML and proposed CL algorithms. We train a vanilla neural network (VNN) model in the CL fashion using deep feature embedding of HTML contents. We compare the proposed CL algorithms with the VNN model trained from scratch and with transfer learning (TL). We show that CL algorithms maintain accuracy over time with a tolerable deterioration of 2.45%. In contrast, VNN and TL-based models' performance deteriorates by over 20.65% and 8%, respectively.