A Novel Consensus Ranking‐Based Ensemble Feature Selection for Efficient Phishing Attack Detection

Phishing attacks continue to represent a serious and evolving threat to individuals, businesses, and the wider field of cybersecurity. Despite advancements in machine learning‐based detection methods, the need to rely on large numbers of features and external data sources leads to various problems, such as greater computational complexity, reduced scalability, and higher latency. Because of these limitations, such systems are not feasible to implement in real world situations. In this study, a novel framework is presented which makes use of lightweight, third party independent features that have been optimized by means of an ensemble feature selection and consensus ranking procedure. The framework is able to efficiently find the most discriminative features by taking advantage of the individual advantages of filter, wrapper, and embedded methods. By using consensus ranking aggregation, this method reduces the biases inherent in each technique and at the same time improves the robustness and generalizability of the feature selection process. The framework was tested using the comprehensive and up to date PhiUSIIL Phishing URL Dataset, which includes more than 235 000 URLs, thus allowing for a thorough evaluation of its performance. We chose the following machine learning models, Random Forest, Decision Tree, Naive Bayes, and Logistic Regression, in order to assess the effectiveness of the best feature set. The reason for selecting these models rather than more advanced deep learning models is that they offer better interpretability, are more computationally efficient, have faster inference speeds, and require fewer resources. Our experiments demonstrated strong performance in detecting phishing URLs, indicating the effectiveness of the ensemble feature selection method and its promise for use in scalable real time detection of malicious websites.