FusionPhishGuard

Phishing detection, mobile and web security, attention mechanism, multi-branch deep learning, word embeddings, transformer, large language models (LLMs), BiLSTM, fusion framework. To address these challenges, we propose FusionPhishGuard, an attention-enhanced multi-branch deep learning framework designed for intelligent phishing detection across mobile and web environments. The framework integrates multiple levels of representation—from raw lexical structures to high-level semantic features—using adaptive attention-based fusion to model complex relationships between URL tokens, subwords, and contextual cues.

FusionPhishGuard

An Attention-Enhanced Multi-Branch Framework for Intelligent Phishing Detection on Mobile and Web Platforms

---

Overview

FusionPhishGuard is an attention-enhanced multi-branch deep learning framework designed for intelligent phishing detection across mobile and web platforms.

The framework integrates:

• Multi-granular tokenization
• Transformer-based embeddings
• Large Language Model embeddings
• Attention-enhanced feature fusion
• Sequential BiLSTM modeling

to effectively detect modern phishing attacks, including:

• Obfuscated phishing URLs
• Brand impersonation attacks
• Mobile redirect phishing
• Cross-platform phishing campaigns
• Semantic deception attacks

---

Publication

📖 Published in IEEE COMSNETS 2026 (SysAI Track)

Paper Title

FusionPhishGuard: An Attention-Enhanced Multi-Branch Framework for Intelligent Phishing Detection on Mobile and Web Platforms

Authors

• Yashwanth Yallavula
• Panigrahi Srikanth
• Manoj Kumar Sunkara
• Vishwanath Tangella

Publication Details

• Conference: IEEE International Conference on Communication Systems & Networks (COMSNETS 2026)
• Track: SysAI (Systems for Artificial Intelligence)
• Publisher: IEEE
• Indexing: Scopus Indexed

Abstract

FusionPhishGuard is an attention-enhanced multi-branch deep learning framework for intelligent phishing detection across mobile and web platforms. The framework integrates multi-granular tokenization, classical word embeddings, transformer encoders, and large language model representations through a gated fusion-attention mechanism. A BiLSTM-based sequential classifier captures contextual dependencies within URLs and web links, enabling robust detection of evolving phishing attacks. Experimental evaluation on the CatchPhish and PhishDump benchmark datasets demonstrates strong performance and generalization across diverse phishing scenarios.

Research Highlights

• Multi-Granular URL Tokenization
• Multi-Branch Embedding Architecture
• Attention-Based Feature Fusion
• BiLSTM Sequential Context Modeling
• Mobile and Web Phishing Detection
• Benchmark Evaluation on CatchPhish and PhishDump
• Published in IEEE COMSNETS 2026
• Scopus Indexed Research Publication

---

Authors

• Yashwanth Yallavula
• Panigrahi Srikanth
• Manoj Kumar Sunkara
• Vishwanath Tangella

---

Key Highlights

• Multi-Branch Deep Learning Framework
• Attention-Based Adaptive Feature Fusion
• Multi-Granular URL Tokenization
• Transformer + LLM Hybrid Representation Learning
• BiLSTM Sequential Context Modeling
• Robust Cross-Platform Phishing Detection
• Strong Performance on Benchmark Datasets
• Interpretable Phishing Feature Learning

---

Framework Architecture

FusionPhishGuard combines heterogeneous embedding paradigms using an adaptive gated attention fusion mechanism.

The framework integrates:

Branch	Purpose
Word2Vec	Lexical pattern learning
FastText	Character-level robustness
BERT	Contextual understanding
RoBERTa	Deep semantic encoding
MiniLM	Lightweight contextual modeling
Qwen	LLM semantic reasoning
Falcon	Large-scale language understanding

The fused representations are processed through:

• Fusion Attention Layer
• Squeeze-and-Excitation Refinement
• BiLSTM Sequential Modeling
• Final Classification Network

---

Experimental Pipeline

The pipeline consists of:

1. URL Collection
2. Multi-Granular Tokenization
3. Embedding Extraction
4. Attention-Based Fusion
5. Sequential Context Modeling
6. Phishing Classification
7. Interpretation & Evaluation

---

Datasets

CatchPhish D2

• Real-world phishing URL dataset
• Includes obfuscated and deceptive URL patterns
• Balanced phishing and legitimate samples

PhishDump

• Large-scale phishing URL corpus
• Mobile and web phishing links
• Real-world phishing campaigns

---

Experimental Results

CatchPhish Results

Model	Accuracy	F1-Score	MCC
Gemma-2 + Best MLP	95.16%	95.19%	0.9044
Fusion (MiniLM + Gemma)	94.88%	94.84%	0.8978
Qwen + MLP	93.96%	93.68%	0.8802
DistilBERT	93.81%	93.71%	0.8761
BERT	93.65%	93.58%	0.8731

---

PhishDump Results

Model	Accuracy	F1-Score	MCC
Qwen + MLP	96.99%	96.43%	0.9386
1D CNN	95.96%	95.17%	0.9175
DistilBERT	95.92%	95.32%	0.9176
BERT	95.74%	94.90%	0.9132
Gemma-2 + MLP	95.46%	94.85%	0.9095

---

FusionPhishGuard employs:

• Gated Additive Attention
• Adaptive Branch Weighting
• Squeeze-and-Excitation Refinement

to dynamically learn the importance of each embedding branch.

---

Ablation Study

Configuration	CatchPhish	PhishDump
Word Embeddings	91.5%	93.8%
+ Transformer Encoders	93.0%	95.2%
+ LLM Branches	95.4%	96.9%
+ Fusion (w/o Attention)	95.2%	97.3%
+ Full Attention Fusion	96.9%	98.1%

---

Repository Structure

FusionPhishGuard/
├── assets/
├── configs/
├── datasets/
├── notebooks/
├── src/
├── results/
├── pretrained_models/
├── README.md
└── requirements.txt

---

Installation

Clone Repository

git clone https://github.com/yourusername/FusionPhishGuard.git
cd FusionPhishGuard

---

Create Environment

conda create -n fusionphish python=3.10
conda activate fusionphish

---

Install Dependencies

pip install -r requirements.txt

---

Training

Train Transformer Models

python src/training/train_transformers.py

Train Embedding Models

python src/training/train_embeddings.py

Train Final Fusion Model

python src/training/train_fusion.py

---

Inference

python src/inference/predict.py

---

Evaluation Metrics

The framework is evaluated using:

• Accuracy
• Precision
• Recall
• F1-Score
• Matthews Correlation Coefficient (MCC)

---

Research Contributions

• Unified mobile + web phishing framework
• Attention-enhanced multi-branch fusion
• Integration of Transformer and LLM embeddings
• Sequential URL dependency modeling
• Robust phishing generalization capability
• Strong benchmark performance

---

Citation

@inproceedings{fusionphishguard2026,
  title={FusionPhishGuard: An Attention-Enhanced Multi-Branch Framework for Intelligent Phishing Detection on Mobile and Web Platforms},
  author={Yallavula Yashwanth, Srikanth Panigrahi, Manoj Kumar Sunkara, Tangella Vishwanath},
  booktitle={IEEE COMSNETS 2026},
  year={2026}
}

Acknowledgements

The authors would like to express their sincere gratitude to Dr. Srinivasa Rao Routhu for his valuable contributions to phishing detection research and for providing access to the benchmark datasets that enabled the experimental evaluation presented in this work.

We also acknowledge:

• IEEE COMSNETS 2026 (SysAI Track)
• CatchPhish Dataset
• PhishDump Dataset
• PyTorch Ecosystem
• Hugging Face Transformers
• Open-source Cybersecurity Research Community

---

FusionPhishGuard — Advancing Intelligent Phishing Detection with Attention-Enhanced Multi-Branch Learning