A deep learning framework for SMS spam detection: Comparative performance of LSTM and BiLSTM Models

Short Message Service (SMS) continues to be an important communication medium for personal, commercial, and institutional use because of its simplicity, accessibility, low cost, and immediate delivery. Despite the rapid growth of internet-based messaging platforms, SMS remains widely used for banking alerts, authentication codes, service notifications, promotional communication, and direct personal interaction. However, the increasing spread of unsolicited and deceptive SMS messages has made spam detection a critical challenge in modern mobile communication. Spam messages often include fraudulent offers, phishing links, fake alerts, and misleading promotional content, which may reduce user trust and create privacy and security risks. Conventional filtering approaches based on rules, blacklists, and manually selected keywords often struggle to adapt to the changing linguistic patterns and structural variations of spam messages. In this context, deep learning provides a more adaptive and effective solution for short text classification.
This study proposes a deep learning-based SMS spam detection framework using Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (BiLSTM) models. For experimental evaluation, a synthetically prepared and labelled SMS dataset was used. Before model training, the dataset was pre-processed through lowercase conversion, URL removal, number removal, punctuation removal, tokenization, and sequence padding. These preprocessing steps transformed raw SMS text into a structured numerical format suitable for sequence-based learning. Both models were implemented and evaluated under the same experimental settings to ensure a fair comparison. The framework used a vocabulary size of 20,000 words, a sequence length of 100, and an 80:20 train-test split. Model performance was evaluated using standard classification metrics, including accuracy, precision, recall, and F1-score.
The experimental results showed highly effective classification performance on the prepared dataset, with both the LSTM and BiLSTM models achieving 100.00% test accuracy in the reported experiment. These findings indicate that sequence-based deep learning architectures can be highly effective for SMS spam detection within the experimental setting of this study. However, the reported performance should be interpreted carefully, as it may also reflect the characteristics and separability of the prepared dataset. Overall, this study provides a comparative evaluation of LSTM and BiLSTM models and offers a useful foundation for future research on intelligent message filtering, secure mobile communication, and more robust spam detection systems.