Establish multilingual generalisation of AURA

Establish whether AURA’s URL feature pipeline and DistilBERT-LoRA encoder generalise reliably to non-English phishing campaigns, including across diverse language families.

Background

AURA is trained and evaluated exclusively on English-language email datasets. The paper therefore identifies uncertainty about whether its URL representations and English-focused DistilBERT-LoRA semantic encoder can detect phishing campaigns expressed in other languages and regional contexts.

Resolving this problem would require multilingual evaluation and likely multilingual training data, potentially using a multilingual DistilBERT backbone and URL corpora covering diverse language families.

References

All constituent datasets comprise English-language content exclusively, and generalisation of AURA's URL feature pipeline and DistilBERT-LoRA encoder to non-English campaigns cannot be assumed.

— AURA: Adaptive Uncertainty-Routed Analysis for Email Threat Detection  (2609.19873 - Berjawi et al., 17 Sep 2026) in Section 7.2, “Language Coverage”