A Lightweight Hybrid BiLSTM–AdaBoost–SVM Framework for Imbalanced SMS Spam Detection

Abstract

SMS-based communication remains a primary vector for smishing attacks and unsolicited spam, posing significant financial and security risks to mobile users. Existing detection methods struggle with two persistent challenges: severe class imbalance (spam constitutes only ≈13% of messages) and the inability of sparse lexical representations to capture informal, abbreviated SMS language. This paper proposes a lightweight hybrid framework—BiLSTM–AdaBoost–SVM—designed for accurate minority-class spam detection without the computational overhead of large transformer-based models. A Bidirectional LSTM encoder first extracts contextual sequence representations; an AdaBoost ensemble produces class-probability outputs that are concatenated with the BiLSTM features to form an enriched 258-dimensional representation; finally, an RBF-kernel SVM learns a maximum-margin decision boundary over this augmented space. Evaluated on the UCI SMS Spam Collection dataset across five independent random seeds, the proposed framework achieves a mean accuracy of 98.03% ± 0.31%, spam precision of 95.68% ± 0.82%, spam recall of 89.26% ± 1.14%, and spam F1-score of 92.36% ± 0.67%—substantially outperforming TF–IDF baselines in recall (+7 to +16 pp, p < 0.01) and achieving a superior precision–recall trade-off for rare-event spam detection, at a fraction of the computational cost of transformer-based models.