Deep Learning Based Real Time Phishing Website Detection System Using URL Features

Abstract

One of the most common and detrimental cyber threats in today's digital world is still phishing websites. 
The threat actors using phishing techniques are simply faking the URLs of the trusted web services and 
getting the users' most sensitive information. Traditional methods of phishing detection have mostly been 
blacklisting and rule-based techniques. However, these methods are limited in facing dynamic and zero
day phishing attacks due to their static characteristics. Researchers responded to these limitations by 
carrying out a thorough experimental study of a deep learning technique a powerful tool for real time 
detection of phishing sites where only the attributes within the URL are the sources. The presented 
mechanism extracts the lexical, structural, and statistical properties of URLs and does not depend on 
other sources such as webpage content analysis or external queries. The system takes advantage of a 
noise-eliminating preprocessing pipeline consisting of feature cleaning, normalization, and class 
imbalance handling through SMOTE. The process concludes with the design and optimization of a deep 
neural network for binary classification. The results of the experiments on the unseen dataset reveal that 
the system gets 95.07% accuracy, 95.69% precision, 96.38% recall, 95.03% F1-score, and AUC of 
99.62%, which shows that it can detect effectively and also has good generalization. Besides, the system's 
inference latency is kept at a low level, thus making it appropriate for real-time use. A comparative 
evaluation indicates that the deep learning-based URL analysis is way better than the conventional 
machine learning techniques, thus providing a detection mechanism that is both scalable and efficient 
for practical phishing website detection.