Abstract:

Phishing attacks are a major cybersecurity threat, often relying on misleading URLs to deceive users into sharing confidential information. This research introduces an innovative hybrid approach for detecting phishing URLs, combining deep learning techniques for extracting temporal features with advanced metaheuristic optimization strategies. The proposed method initiates with a robust multi-stage feature selection strategy, which includes mutual information filtering, recursive feature elimination, model-driven importance ranking, and pruning based on feature correlation. This ensures that only the most significant and non-overlapping features are chosen for the next stages of the learning process. A Bidirectional Long Short- Term Memory (BiLSTM) network is then applied to extract deep temporal representations from the preprocessed feature vectors. These temporal embeddings are subsequently fed into a Support Vector Machine (SVM) classifier to perform the final prediction. To improve classification accuracy, we introduce a hybrid metaheuristic approach hEVOSWO that simultaneously optimizes feature selection and hyperparameters by combining Evolutionary Optimization (EVO) and SpiderWasp Optimization (SWO) which enables efficient exploration and fine-tuning within
the search space of SVM parameter settings. Experimental results on the benchmark ISCX-URL2016 dataset demonstrate that our proposed approach achieves an accuracy of 97.31%, a precision of 97.42%, a recall of 97.18%, and an F1-score of 97.30%. The findings validate the efficiency and robustness of the proposed framework, highlighting its potential for advancing resilient phishing detection systems.