International Journal of Advanced Research in Computer and Communication Engineering

Abstract: Phishing and smishing attacks have rapidly increased in digital communication platforms, exploiting user trust to steal personal and financial information. Traditional blacklist and rule-based detection systems lack adaptability and fail to detect evolving or zero-day attacks. Although deep learning has shown promise in text-based threat detection, existing research often focuses on a single architecture, lacks real-world datasets, or provides limited benchmarking across models. To address these gaps, this study presents a comparative evaluation of three deep learning models—RNN-LSTM, RNN-GRU, and GloVe-enhanced LSTM—for phishing and smishing text classification. A dataset of 27,000 real-world messages, collected from cybersecurity units and extended with controlled synthetic samples, was preprocessed using tokenization, stemming, padding, and semantic embeddings. Each model underwent structured hyperparameter optimization with dropout, L2 regularization, and early stopping to enhance generalization. Experimental results show that the GloVe-LSTM model achieved the highest performance with 90.07% test accuracy and a 90.16% F1-score, closely followed by tuned LSTM and GRU models. Statistical validation using a McNemar test confirmed no significant difference in model performance (p > 0.05). These findings demonstrate that semantic embeddings significantly improve phishing and smishing detection accuracy, supporting scalable deployment in cybersecurity systems such as email filtering, telecom SMS gateways, and digital fraud prevention platforms.

Keywords: Phishing Detection; Smishing; Deep Learning; LSTM; GRU; GloVe Embeddings; NLP; Cybersecurity

Downloads: | DOI: 10.17148/IJARCCE.2025.1411149