Detection Model for URL Phishing with Comparison Between Shallow Machine Learning and Deep Learning Models
Nizam Aditya Zuhayr, Girinoto, Nurul Qomariasih, Hermawan Setiawan · Advances in intelligent systems research/Advances in Intelligent Systems Research · 2023
In the report on trends in phishing activity released by the Anti-Phishing Working Group (APWG), global phishing cases continued to increase throughout 2021 to the first quarter of 2022.This study compares shallow machine learning algorithms that have been used by governments with deep learning in classifying URLs.Phishing.From the data as many as 30,047 URLs consisting of 15,022 phishing URLs and 15,025 legal URLs, the distribution was carried out for training data and test data.URL phishing modeling uses deep learning algorithms LSTM and GRU as well as the best shallow machine learning algorithms from research conducted by Rao et.al, namely Random Forest (RF), Logistic Regression (LR), and Decision Tree (DT).Modeling is done based on URL characteristics, text structure, and a combination of URL characteristics with text structure.Based on URL characteristics, the model with the best accuracy from the shallow machine learning algorithm is Random Forest at 97.4%, while the deep learning algorithm is LSTM at 96.7%.Based on the structure of the text, the best deep learning algorithm is the GRU of 97.8%.While the combination model using 2 deep learning algorithms LSTM and GRU get an accuracy of 98.1%.Furthermore, the combination model as the best model is implemented in the form of a website using the Flask framework with the classification results in the form of a URL probability score that is detected as a phishing URL.