Focal Loss and Self-Attention Convolutional Neural Network for Phishing Detection Technique

In real life there are typically more legitimate URLs than phishing ones, conventional machine learning and deep learning algorithms tend to misclassify phishing URLs and favour benign ones. Few researches have been conducted to address the imbalance problem in phishing detection datasets, which severely compromises companies' financial data and digital privacy. To mitigate the impact of imbalance and enhance the detection performance of phishing emails, this paper proposes an improved Convolutional Neural Network (CNN) using focal Loss function, which assign more weight to samples that are difficult to classify than easier to classify. Moreover, the self-attention network (SAN) enhanced the CNN classification performance. This mechanism allows the network to focus on small instances of the feature graph and capture deep features by assigning different attentional weights to different parts of the input. As a result, the network has become stronger and able to understand the relationships between the inputs. Experimental results have demonstrated that the proposed algorithms outperform other deep learning algorithms, achieving an accuracy of 99.5%, precision of 99.63%, Sensitivity of 99.5% and F1-score of 99.78%.