Image Based Phishing Detection Using Transfer Learning

Abstract

This paper considers the problem of a phishing detection based on the visual appearance of a login page of a website. Two main procedures which are the login page data augmentation and the login page classification are taken into account to achieve the phishing detection task. Convolutional neural networks (CNNs) are applied to the classification task since they justify their effectiveness on the image classification during the current decade but training the whole weights of a CNN model requires a large dataset and spends a lot of time. This leads to the requirement of the data augmentation to produce more data from the original data. This paper proposes a subimage random placement method for the data augmentation task which is the main contribution of this paper. Transfer learning is also applied to reduce a long training time by using a pre-trained model and training only the fully connected layer for a new classification task. This paper will conduct the experiment in two aspects, the binary classification and the multiclass classification, and the accuracy comparisons among various pre-trained CNNs. The result will show that transfer learning works well with the data that are generated in the same way of the proposed data augmentation.