Improving Detection Accuracy of Phishing Attacks with Feature Selection and Machine learning Techniques in Cybersecurity

Phishing attacks are a very serious issue that keeps on emerging in the online world with cybercriminals continuously finding new tricks to deceive users and steal valuable information. This study aims at solving the urgent issue of phishing identification with the help of the latest methods of ML capabilities. This study focuses on using state-of-the-art ML approaches to tackle the urgent issue of phishing detection. This study presents a robust detection model for phishing URLs using a Bidirectional Long Short-Term Memory (BiLSTM) model, which can efficiently handle high data volumes. The study combines the exploratory data analysis and feature-importance analysis to recognize the essential URL features, whereas Min-Max scaling and SMOTE are employed to guarantee the existence of balanced and substantial learning. The proposed BiLSTM identifies the forward and backward contextual dependencies in URL sequences, improving its capacity to identify complex phishing patterns. The experimental findings indicate that the proposed model has 98.14% accuracy, 98.70 precision, 98.39% recall, 98.25 F1-score and AUC of 0.9997. The BiLSTM has been compared with both traditional ML models (DecisionTree, NaiveBayes, Random Forest, Logistic Regression, AdaBoost) and advanced DL models (LSTM, CNN-LSTM) and can be stated as always being better than these methods. The findings indicate that bidirectional sequence learning is more effective in learning patterns of URLs than in a unidirectional or shallow model, and results in improved generalization, less overfitting, and more reliability. Generally, the BiLSTM model creates a strong and scalable phishing detection model, which has enhanced the position of cybersecurity solutions.