Cross-lingual SMS spam detection using GAN-based augmentation for imbalanced datasets

SMS spam detection remains a critical challenge in mobile communication security, particularly when addressing the inherent class imbalance present in real-world datasets, where spam messages constitute only 13–15% of total communications. This study presents a comprehensive framework integrating advanced word embeddings, deep learning architectures, and Generative Adversarial Networks (GANs) for synthetic data augmentation to enhance SMS spam classification performance. A systematic evaluation is conducted across six machine learning algorithms (Support Vector Machine (SVM), Logistic Regression (LR), K-Nearest Neighbors (KNN), Decision Tree (DT), Stochastic Gradient Descent (SGD), Random Forest (RF)) and two deep learning models (Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM)), combined with five embedding techniques (Term Frequency–Inverse Document Frequency (TF-IDF), Bag of Words (BoW), Word2Vec, GloVe, Bidirectional Encoder Representations from Transformers (BERT)), resulting in 120 experimental configurations tested both with and without data augmentation. A novel GAN-based approach is employed to generate synthetic word embeddings rather than raw text, preserving semantic coherence while addressing dataset imbalance more effectively than traditional oversampling methods (Synthetic Minority Over-sampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN)). Experimental validation on both the monolingual UCI SMS Spam Collection and multilingual datasets demonstrates that the optimal BERT+Bi-LSTM+GAN configuration achieves exceptional performance, with F1-scores of 97.61% (monolingual) and 94.44% (multilingual), surpassing existing state-of-the-art approaches. The comprehensive evaluation framework, incorporating Matthews Correlation Coefficient (MCC) and Cohen’s Kappa (CK), provides robust assessment for imbalanced classification scenarios. Results reveal that contextual embeddings consistently outperform traditional frequency-based methods, with BERT achieving perfect precision of 100% in baseline configurations. The study establishes strategic deployment guidelines: BERT configurations for maximum accuracy scenarios, Word2Vec approaches for balanced performance–efficiency requirements, and traditional methods for resource-constrained environments. Cross-linguistic validation confirms the universality of the approach, demonstrating only a 3.25% performance degradation in multilingual contexts. This research advances both theoretical understanding of imbalanced text classification and practical implementation of robust SMS spam detection systems, providing a methodological foundation applicable to broader cybersecurity and natural language processing challenges.