Self-Training Approach for Smishing Detection in Multilingual and Low-Resource Settings

Smishing--phishing conducted via SMS--continues to spread globally. Most detection systems are built using only English data, which limits their use for other languages. We propose a multilingual smishing detection framework based on self-training with XLM-RoBERTa. Starting with English-labeled data, we apply zero-shot inference to Bengali and Swahili messages, extract high-confidence predictions, and incorporate them as pseudo-labeled examples for fine-tuning. Our approach improves recall and F1 scores in both target languages without requiring parallel corpora or costly annotations. We further analyze the impact of language-specific and balanced pseudo-label augmentation through ablation studies. The results show that the combination of samples from multiple languages leads to better generalization and reliability. This work highlights an effective low-resource strategy for building multilingual smishing detectors, allowing a broader deployment across linguistically diverse user populations.