With rapid internet advancements, the impact of phishing attacks, frequency, and severity are escalating. Phishing attempts to mimic the websites of official businesses, including financial institutions, banks, government offices, and e-commerce platforms. Phishing attacks on websites target at tricking users of the internet into disclosing personal data, such as financial data, login credentials. Accordingly, enhancing an effective phishing diagnosis system is important to guarantee cryptocurrency transactions’ security and reliability. The practical path in diagnosing phishing attacks is applying ML methods. For decreasing phishing attacks’ diagnosis error, a 2-stage strategy is shown here. In this paper, we present a novel hybrid framework for improving phishing detection by including advanced feature preprocessing, hybrid optimization, and various attention mechanisms into deep learning models. Our method uses two benchmark datasets (PhiUSIIL and Tan) to compare the performance of LSTM, CNN, and MLP architectures. Our method uses two benchmark datasets (PhiUSIIL and Tan) to compare the performance of LSTM, CNN, and MLP architectures. We create an upgraded Hunger Games Search (HGS) algorithm to extract the most discriminating elements and integrate them with CNN-based visual representations. To improve model interpretability and focus, we use Self-Attention for LSTM (to capture sequential dependencies), Channel Attention in CNN (to detect phishing-relevant patterns in picture representations), and Channel Attention in MLP (for feature refinement). A hybrid technique is used to tune hyperparameters, combining Harris Hawks Optimization (HHO) and the Tree-structured Parzen Estimator (TPE), allowing for more robust model generalization. Our technique outperforms baseline models, achieving an accuracy of 99.95% on the PhiUSIIL dataset and 97.00% on the Tan dataset. Evaluation criteria such as F1-Score, Precision, Recall, and ROC-AUC demonstrate the suggested method's usefulness and generalizability. Our experimental results indicate significant improvements in phishing diagnostic accuracy, highlighting the efficacy of several strategies in real-world apps.
