Earticle

Phishing attacks increasingly exploit mobile platforms and target users communicating in lowresource or code-mixed languages, posing challenges to traditional centralized detection systems. This study proposes hybrid knowledge distillation and federated learning framework for real-time, on-device phishing detection. The approach integrates a fine-tuned XLM-RoBERTa "teacher" model with a compact MobileBERT "student" model, distilled to achieve near teacher-level performance while enabling efficient offline inference. The distilled MobileBERT model, converted to ONNX for platform portability, achieved a fivefold reduction in size while preserving 98% of the original model’s accuracy. We conducted zero-shot evaluations on Korean, Spanish, and Turkish datasets to guarantee crosslinguistic robustness, and consistently obtained good accuracy and recall rates. Moreover, privacy-preserving updates were made possible using a federated learning simulation, which permits decentralized model enhancements without data exchange. The suggested architecture offers a cost-effective, expandable, and privacy-conscious approach to phishing detection in scenarios with limited connection and linguistic diversity.

This paper measures the end-to-end impact of inserting a Jejueo→Standard-Korean normalization layer before a retrieval-augmented generation (RAG) QA system and of merging results from two retrieval paths. Under an identical stack, we compare: (i) the original-Jejueo retrieval path, retrieve with the unnormalized query (RAW), (ii) the normalized-to-Standard-Korean retrieval path, retrieve with the query rewritten by our Jejueo→Standard-Korean normalizer (NORM), and (iii) the merged path, deduplicate and re-rank the union of RAW and NORM candidates (DUAL). Using the Jejueo Interview Transcripts, lightweight preprocessing, and a compact LoRA-based normalizer, we evaluate retrieval/ranking (Hit@5, MRR@5), evidence attribution (Attrib-F1), answer quality (Token-F1/ROUGE-L), diversity (ILD@5), and hallucination. Normalization alone improves top-rank concentration and citation precision over RAW (Token-F1 +6–7; Attrib-F1 +0.09; Hit@5 ≈ +0.11), and DUAL adds further gains (≈ +4 Token-F1; +0.06 Attrib-F1), increases diversity, and lowers hallucination (≈ −1.7%p). Error analyses trace benefits to aligning Jejueo sentence-final endings, discourse markers, and idioms that otherwise destabilize retrieval; DUAL also hedges over/under-substitution from imperfect normalization. Our intent is practical: use DUAL by default for robustness, and prefer NORM alone when normalization confidence is high and latency/compute budgets are tight.