A Deep Learning Framework for Phishing URL Detection Using Multilevel Embeddings and Attention Based Feature Fusion

Phishing attacks remain a major cybersecurity threat, while conventional blacklist- and handcrafted feature-based detection methods have limited capability to identify previously unseen and rapidly evolving phishing URLs. This study proposes PhishDetect, a representation-aware multi-branch deep learning framework that learns complementary structural and contextual characteristics directly from URLs. Unlike approaches that rely on a single representation level, PhishDetect constructs character-level, word-level, and character–word-level embeddings and maps them to specialized Char-CNN, LSTM, and Bi-LSTM branches, respectively. The Char-CNN captures fine-grained lexical and structural patterns, the LSTM models sequential dependencies among URL tokens, and the Bi-LSTM learns bidirectional contextual relationships from the combined character–word representation. The heterogeneous branch features are subsequently integrated into a unified representation and processed by an attention-based feature weighting mechanism that emphasizes discriminative information before classification. Experiments were conducted using the public Phishing Site URLs dataset, from which a balanced subset of 120,000 URLs was used for model development and evaluation. Feature-dimensionality analysis identified a 155-dimensional fused representation as the most effective configuration. The proposed framework achieved 97.06% accuracy, 96.75% precision, 97.38% recall, and a 97.07% F1-score. Component-level ablation experiments further demonstrated the complementary contribution of the multi-branch architecture and attention mechanism. These findings indicate that representation-aware multi-level learning and adaptive feature weighting provide an effective approach for phishing URL detection.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11263678/v1
Primary Topic
Spam and Phishing Detection
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Deep Learning Framework for Phishing URL Detection Using Multilevel Embeddings and Attention Based Feature Fusion

Yikai Liu, Alok Kumar Sharma, Chuan-Bi Lin
4 links
Research Square
Spam and Phishing Detection
preprint

A Deep Learning Framework for Phishing URL Detection Using Multilevel Embeddings and Attention Based Feature Fusion

Yikai Liu, Alok Kumar Sharma, Chuan-Bi Lin
preprint en

Abstract

Phishing attacks remain a major cybersecurity threat, while conventional blacklist- and handcrafted feature-based detection methods have limited capability to identify previously unseen and rapidly evolving phishing URLs. This study proposes PhishDetect, a representation-aware multi-branch deep learning framework that learns complementary structural and contextual characteristics directly from URLs. Unlike approaches that rely on a single representation level, PhishDetect constructs character-level, word-level, and character–word-level embeddings and maps them to specialized Char-CNN, LSTM, and Bi-LSTM branches, respectively. The Char-CNN captures fine-grained lexical and structural patterns, the LSTM models sequential dependencies among URL tokens, and the Bi-LSTM learns bidirectional contextual relationships from the combined character–word representation. The heterogeneous branch features are subsequently integrated into a unified representation and processed by an attention-based feature weighting mechanism that emphasizes discriminative information before classification. Experiments were conducted using the public Phishing Site URLs dataset, from which a balanced subset of 120,000 URLs was used for model development and evaluation. Feature-dimensionality analysis identified a 155-dimensional fused representation as the most effective configuration. The proposed framework achieved 97.06% accuracy, 96.75% precision, 97.38% recall, and a 97.07% F1-score. Component-level ablation experiments further demonstrated the complementary contribution of the multi-branch architecture and attention mechanism. These findings indicate that representation-aware multi-level learning and adaptive feature weighting provide an effective approach for phishing URL detection.

Research Square
Chaoyang University of Technology (TW)
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