Context-less lightweight phishing domain detection for mobile and edge devices via INT8-quantized low-parameter transformer classifiers

Abstract Mobile devices constitute the front door for billions of users to gain access to the internet, but are also extremely susceptible to phishing attacks launched via adversarial domain creation. The current transformer based Domain classifiers provide high accuracy, but have a large number of parameters more than 86.9 M, which is too large for edge devices with memory and battery constraints. We propose a context-less phishing detection framework which classifies domains based on the analysis of only the domain string itself, avoiding the risk of downloading malicious page content, and saving network bandwidth. All 4 variants of the lightweight transformer classifiers: PhishNet-10M, PhishNet-20M, PhishNet-45M and PhishNet-85M are trained and evaluated, with INT8 post training quantization for maximum edge efficiency. Validation with early stopping up to 20 epochs of training shows that the INT8 quantized version of the PhishNet-10M model achieves a validation accuracy of 95.72% and PhishNet-85M model achieves a validation accuracy of 96.20%. The 10.8 M (PhishNet-10M) configuration represents an ideal point of deployment. We benchmark all PhishNet variants against established sequence models (LSTM, GRU, CNN) and strong distilled transformer baselines (TinyBERT), demonstrating superior accuracy and efficiency for resource-constrained mobile and IoT edge devices.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-12
DOI
https://doi.org/10.1038/s41598-026-74278-w
Primary Topic
Spam and Phishing Detection
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Context-less lightweight phishing domain detection for mobile and edge devices via INT8-quantized low-parameter transformer classifiers

Arshpreet Kaur, Achin Jain, Arun Kumar Dubey, Yash Prakash Narayan et al.
Scientific Reports
Spam and Phishing Detection
article

Context-less lightweight phishing domain detection for mobile and edge devices via INT8-quantized low-parameter transformer classifiers

Arshpreet Kaur, Achin Jain, Arun Kumar Dubey, Yash Prakash Narayan, Deepak Vishwakarma
article en

Abstract

Abstract Mobile devices constitute the front door for billions of users to gain access to the internet, but are also extremely susceptible to phishing attacks launched via adversarial domain creation. The current transformer based Domain classifiers provide high accuracy, but have a large number of parameters more than 86.9 M, which is too large for edge devices with memory and battery constraints. We propose a context-less phishing detection framework which classifies domains based on the analysis of only the domain string itself, avoiding the risk of downloading malicious page content, and saving network bandwidth. All 4 variants of the lightweight transformer classifiers: PhishNet-10M, PhishNet-20M, PhishNet-45M and PhishNet-85M are trained and evaluated, with INT8 post training quantization for maximum edge efficiency. Validation with early stopping up to 20 epochs of training shows that the INT8 quantized version of the PhishNet-10M model achieves a validation accuracy of 95.72% and PhishNet-85M model achieves a validation accuracy of 96.20%. The 10.8 M (PhishNet-10M) configuration represents an ideal point of deployment. We benchmark all PhishNet variants against established sequence models (LSTM, GRU, CNN) and strong distilled transformer baselines (TinyBERT), demonstrating superior accuracy and efficiency for resource-constrained mobile and IoT edge devices.

Scientific Reports
INTI International University (MY), Defence Research and Development Organisation (IN), Bharati Vidyapeeth's College of Engineering, Delhi (IN), Manipal University Jaipur (IN)
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