View in EDS

A compound-aware encoder-only transformer model for smishing detection.

Saved in:
Bibliographic Details
Title: A compound-aware encoder-only transformer model for smishing detection.
Authors: Nguyen-Xuan, Sam1 (AUTHOR) samnx2@fe.edu.vn, Nguyen, Han2 (AUTHOR) han316@usf.edu
Source: Journal of Supercomputing. Mar2026, Vol. 82 Issue 4, p1-25. 25p.
Subjects: Transformer models, Mobile communication system security, Machine learning, Text mining, Encoding, Phishing
Abstract: Smishing attacks increasingly exploit lexical obfuscation and compound-word constructions within both SMS content and embedded URLs, which significantly degrades the effectiveness of conventional tokenization-based detection approaches. To address this challenge, we propose the Compound-Aware Encoder-Only Transformer (CAEoT), a framework that explicitly reconstructs semantically meaningful sub-components from obfuscated compound expressions prior to Transformer encoding. CAEoT introduces a case factory that adaptively activates text- and URL-specific compound-aware decomposition modules, thereby exposing latent semantic cues before tokenization. The resulting enriched token sequences are then processed by an encoder-only Transformer and a lightweight classification head. Experiments conducted on two benchmark smishing datasets demonstrate that CAEoT achieves strong and consistent detection performance, with F1-scores of 0.89 and 0.98, ROC-AUC values of up to 1.00, and PR-AUC values of up to 0.99. Further analysis of ROC and precision–recall curves confirms that CAEoT maintains robust discriminative capability under class imbalance while preserving practical end-to-end inference efficiency. These results indicate that explicit compound-aware decomposition provides a principled and effective enhancement for Transformer-based smishing detection without modifying the underlying encoder architecture. [ABSTRACT FROM AUTHOR]
Copyright of Journal of Supercomputing is the property of Springer Nature and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:Smishing attacks increasingly exploit lexical obfuscation and compound-word constructions within both SMS content and embedded URLs, which significantly degrades the effectiveness of conventional tokenization-based detection approaches. To address this challenge, we propose the Compound-Aware Encoder-Only Transformer (CAEoT), a framework that explicitly reconstructs semantically meaningful sub-components from obfuscated compound expressions prior to Transformer encoding. CAEoT introduces a case factory that adaptively activates text- and URL-specific compound-aware decomposition modules, thereby exposing latent semantic cues before tokenization. The resulting enriched token sequences are then processed by an encoder-only Transformer and a lightweight classification head. Experiments conducted on two benchmark smishing datasets demonstrate that CAEoT achieves strong and consistent detection performance, with F1-scores of 0.89 and 0.98, ROC-AUC values of up to 1.00, and PR-AUC values of up to 0.99. Further analysis of ROC and precision–recall curves confirms that CAEoT maintains robust discriminative capability under class imbalance while preserving practical end-to-end inference efficiency. These results indicate that explicit compound-aware decomposition provides a principled and effective enhancement for Transformer-based smishing detection without modifying the underlying encoder architecture. [ABSTRACT FROM AUTHOR]
ISSN:09208542
DOI:10.1007/s11227-026-08321-y