CSqL-m-H2LA-GRBiTM: categorical square loss-based multi-head hierarchical location attention enabled deep learning mechanism for smishing detection

Smishing detection is an effective research area that remains essential in the recent decade. Detection of SMS phishing is decisive due to adversaries evolving in the network simultaneously. Several existing researches worked to perform accurate smishing detection but failed due to certain challenges. Shortage of large datasets, improper classification of legitimate and malicious messages, high computational resource utilization, and high false positive rates were the major challenges addressed in the conventional researches. Thus, the Categorical Square Loss-based multi-head Hierarchical Location Attention-enabled Gated Recurrent Unit with Bidirectional Long Short-Term Memory (CSqL-m-H2LA-GRBiTM) model is proposed in the research that presents robust smishing detection, leveraging the power of artificial intelligence (AI). The model captures information concerning spatial and temporal features, which elevates the efficiency in handling complex patterns in smishing detection. The proposed m-H2LA mechanism focuses on the relevant messages and their features to boost interpretability. Moreover, the deployment of a CSqL loss function enables robust detection by balancing maximum separation with probabilistic calibration and heavily penalizes misclassifications that cross the safety boundary. Overall, the CSqL-m-H2LA-GRBiTM model attains the accuracy, F1-score, and log-loss of 98.36%, 96.84%, and 5.09 with the SMS phishing dataset and 98.43%, 97.41%, and 5.23 with SMS Smishing Dataset at K-fold-10 respectively, which depicts high efficiency in comparison with the traditional methods.