Detection of Phishing URLs by Using Deep Learning Approach and Multiple Features Combinations
Tomas Rasymas, Laurynas Dovydaitis
Abstract
Phishing detection is mostly performed through the usage of blacklists.However, blacklists cannot be exhaustive and lack the ability to detect newly generated phishing URLs.In recent years, increased attention has been given to exploring machine learning techniques in order to improve the universality of phishing URL detectors.This article aims at presenting our results on phishing URLs classification where three different features: lexical features, character level embeddings, and word level embeddings were compared with the view to find an approach that maximizes the ratio of phishing URL detection.In addition, a new deep neural network architecture for that problem was suggested.The said deep neural network consists of combined multiple CNN and LSTM layers.The 94.4% accuracy was achieved by combining character and word level embeddings.