Phishing Detection Based on Multi-Feature Neural Network

Abstract

Phishing detection methods are used to protect Internet users from leaking private information to phishing websites. However, the passive phishing detection method, which is widely used and based on blacklists, has limitations on timeliness and defense against zero-day phishing attacks and active phishing detection models with single-feature can be easily targeted by attackers. It is necessary to design and establish an active phishing detection model with high timeliness and strong adaptability. We propose a phishing detection method based on multi-feature extraction and deep learning technology. The model is constructed of a multilayer perceptron (MLP) for self-defined feature, a convolutional neural network (CNN) for image feature, a recurrent neural network (RNN) for text feature to extract feature vectors, and a classification network to fuse features and make the judgement. Our model’s accuracy achieves 0.9775 and recall reaches up to 0.9901. Experiment results of our model prove superior performance to those of other classification algorithms, demonstrating our model’s ability to deal with complex and changeable phishing detection tasks at this stage.