Catching Multilingual Spam on a CPU: A Hybrid Detector with Review Tiers

I built this detector after a small English model failed on the first Bengali scam message I tried. The message was short, polite, and asked for a bKash payment. The old model saw none of the usual English spam words and scored it as normal. That failure set the direction for the work. The detector reads short messages in English, Bengali, Banglish, Hindi, Hinglish, Spanish, French, Arabic, and mixed script text. It combines character and word n-grams, 100 hand-built scam signals, a frozen multilingual MiniLM encoder, a calibrated stacker, and a small set of high-confidence rules. The corpus has 150,866 messages after exact and near-duplicate removal. On a held-out test set with a 20 percent spam prior, it reaches a PR AUC of 0.9385, an F1 of 0.8315, a precision of 94.9 percent, a recall of 74.0 percent, a false-positive rate of 0.99 percent, and an ECE of 0.0039. The 150-case adversarial suite gives a 97.33 percent strict pass rate, a 99.33 percent soft pass rate, and no false blocks. The INT8 ONNX encoder runs in 7.5 ms for one message, and a warmed single message API call takes about 15 ms on a four-core Intel i3 CPU. I also report the branch ablations, the per-language results, and the failures that changed the design.

Authors

Publication Details

Journal
Figshare
Published
2026-11-17
DOI
https://doi.org/10.6084/m9.figshare.35040670
Primary Topic
Spam and Phishing Detection
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Catching Multilingual Spam on a CPU: A Hybrid Detector with Review Tiers

Md. Toriqul Haque
3 links
Figshare
Spam and Phishing Detection
preprint

Catching Multilingual Spam on a CPU: A Hybrid Detector with Review Tiers

Md. Toriqul Haque
preprint en

Abstract

I built this detector after a small English model failed on the first Bengali scam message I tried. The message was short, polite, and asked for a bKash payment. The old model saw none of the usual English spam words and scored it as normal. That failure set the direction for the work. The detector reads short messages in English, Bengali, Banglish, Hindi, Hinglish, Spanish, French, Arabic, and mixed script text. It combines character and word n-grams, 100 hand-built scam signals, a frozen multilingual MiniLM encoder, a calibrated stacker, and a small set of high-confidence rules. The corpus has 150,866 messages after exact and near-duplicate removal. On a held-out test set with a 20 percent spam prior, it reaches a PR AUC of 0.9385, an F1 of 0.8315, a precision of 94.9 percent, a recall of 74.0 percent, a false-positive rate of 0.99 percent, and an ECE of 0.0039. The 150-case adversarial suite gives a 97.33 percent strict pass rate, a 99.33 percent soft pass rate, and no false blocks. The INT8 ONNX encoder runs in 7.5 ms for one message, and a warmed single message API call takes about 15 ms on a four-core Intel i3 CPU. I also report the branch ablations, the per-language results, and the failures that changed the design.

Figshare
Spam and Phishing Detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon