Multi-Signal Learning Framework for Robust Detection of Visually Deceptive Text
Name
electronics-15-04531-v2.pdf
Size
1009.62 KB
Format
Adobe PDF
Checksum (MD5)
234efa18799aaa79bf2fbde9afa0b7a0
Author(s)
Jovanovic-Hacon, Aleksandar
Luo, Sean-Winston
Cao, Connie
Wanderley, Miguel
Boriel, Hailey
Tenison, Irene
Kim, Hyunwoo
Santos Beserra, Fernando
Ferreira, Mariana Aparecida
Palacios, Rafael
Date Issued
October 3, 2026
Journal
Electronics
Publisher
Multidisciplinary Digital Publishing Institute
Citation
Jovanovic-Hacon, A.; Luo, S.-W.; Cao, C.; Wanderley, M.; Boriel, H.; Tenison, I.; Kim, H.; Santos Beserra, F.; Ferreira, M.A.; Palacios, R.; et al. Multi-Signal Learning Framework for Robust Detection of Visually Deceptive Text. Electronics 2026, 15, 4531.
Version
Final published version
Abstract
Visually deceptive text, such as homoglyph-based and near-duplicate variations, poses a significant challenge in applications ranging from financial systems to online identity verification. Existing approaches rely on either string-based similarity metrics or learned embeddings, but each captures only a subset of the information required for robust spoof detection. In this work, we show that these methods are complementary: embedding-based models capture perceptual similarity but struggle under structural divergence, while string-based methods capture character-level patterns but fail under visually deceptive transformations. Motivated by this observation, we propose a unified multi-signal learning framework that integrates visually aligned text embeddings with string-based features such as edit distance and token overlap through a learned fusion model. Experiments run on a large homoglyph-domain benchmark show that the proposed method consistently outperforms both string-based and embedding-based baselines, with the largest gains in challenging cases where individual methods fail. The approach is motivated by deployment scenarios such as corporate email gateways and DNS or registrar-side monitoring pipelines, where lightweight detection before delivery or registration would be valuable, although additional techniques to deceive domain names should be considered for an efficient implementation.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Cybersecurity at MIT Sloan
Terms of Use
Creative Commons Attribution
Persistent DSpace Link
DOI of Published Version
https://doi.org/10.3390/electronics15194531
