ALBERT-BiLSTM cross-attention network with progressive knowledge distillation for multi-domain SMS spam classification

SMS spam detection remains a critical challenge in digital communication systems, particularly in agriculture where farmers depend on SMS services for weather updates, crop prices, and government notifications. Traditional spam detection methods fail to handle evolving spam tactics and class imbalan...

Full description

Saved in:
Bibliographic Details
Main Authors: B.S. Aparna, Remya S, Manu J. Pillai, Somula Rama Subbareddy, Yong Yun Cho
Format: Article
Language:English
Published: Elsevier 2025-09-01
Series:Results in Engineering
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S259012302502794X
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849226912323338240
author B.S. Aparna
Remya S
Manu J. Pillai
Somula Rama Subbareddy
Yong Yun Cho
author_facet B.S. Aparna
Remya S
Manu J. Pillai
Somula Rama Subbareddy
Yong Yun Cho
author_sort B.S. Aparna
collection DOAJ
description SMS spam detection remains a critical challenge in digital communication systems, particularly in agriculture where farmers depend on SMS services for weather updates, crop prices, and government notifications. Traditional spam detection methods fail to handle evolving spam tactics and class imbalance effectively. This paper proposes a hybrid ensemble model that combines ALBERT transformer embeddings with BiLSTM networks enhanced by attention mechanisms and knowledge distillation. Experimental evaluation on the SMS Spam Collection dataset demonstrates superior performance with 98% accuracy, 0.98 precision, 0.97 recall, and 0.98 F1-score. The ensemble model significantly outperforms individual ALBERT (88% accuracy) and BiLSTM (54% accuracy) models. Knowledge distillation reduces model size from 425 MB to 67 MB, enabling real-time deployment with 0.045-second processing time per message. The system achieves 22.2 messages per second throughput, making it suitable for practical applications. Real-world testing in agricultural communication networks demonstrates significant improvements in spam filtering compared to traditional keyword-based filters. The hybrid approach provides an effective solution for SMS spam detection with demonstrated real-time performance and practical applicability in resource-constrained environments.
format Article
id doaj-art-0fffc5ff241f40afbabd865fad93434e
institution Kabale University
issn 2590-1230
language English
publishDate 2025-09-01
publisher Elsevier
record_format Article
series Results in Engineering
spelling doaj-art-0fffc5ff241f40afbabd865fad93434e2025-08-24T05:14:20ZengElsevierResults in Engineering2590-12302025-09-012710672710.1016/j.rineng.2025.106727ALBERT-BiLSTM cross-attention network with progressive knowledge distillation for multi-domain SMS spam classificationB.S. Aparna0Remya S1Manu J. Pillai2Somula Rama Subbareddy3Yong Yun Cho4Department of CSE, TKM College of Engineering, Kollam, 691005, Kerala, IndiaDepartment of CSE, Amrita School of Computing, Amrita Vishwa Vidyapeetham, Amritapuri, 690525, Kerala, IndiaDepartment of CSE, TKM College of Engineering, Kollam, 691005, Kerala, IndiaDepartment of Computer Science and Engineering, Symbiosis Institute of Technology, Hyderabad Campus, Symbiosis International (Deemed University), 509217, Pune, IndiaDepartment of Information and Communication Engineering, Sunchon National University, Jeollanam-do, Suncheon, South Korea; Corresponding author.SMS spam detection remains a critical challenge in digital communication systems, particularly in agriculture where farmers depend on SMS services for weather updates, crop prices, and government notifications. Traditional spam detection methods fail to handle evolving spam tactics and class imbalance effectively. This paper proposes a hybrid ensemble model that combines ALBERT transformer embeddings with BiLSTM networks enhanced by attention mechanisms and knowledge distillation. Experimental evaluation on the SMS Spam Collection dataset demonstrates superior performance with 98% accuracy, 0.98 precision, 0.97 recall, and 0.98 F1-score. The ensemble model significantly outperforms individual ALBERT (88% accuracy) and BiLSTM (54% accuracy) models. Knowledge distillation reduces model size from 425 MB to 67 MB, enabling real-time deployment with 0.045-second processing time per message. The system achieves 22.2 messages per second throughput, making it suitable for practical applications. Real-world testing in agricultural communication networks demonstrates significant improvements in spam filtering compared to traditional keyword-based filters. The hybrid approach provides an effective solution for SMS spam detection with demonstrated real-time performance and practical applicability in resource-constrained environments.http://www.sciencedirect.com/science/article/pii/S259012302502794XSpam detectionBiLSTMALBERTKnowledge distillationDeep learningText classification
spellingShingle B.S. Aparna
Remya S
Manu J. Pillai
Somula Rama Subbareddy
Yong Yun Cho
ALBERT-BiLSTM cross-attention network with progressive knowledge distillation for multi-domain SMS spam classification
Results in Engineering
Spam detection
BiLSTM
ALBERT
Knowledge distillation
Deep learning
Text classification
title ALBERT-BiLSTM cross-attention network with progressive knowledge distillation for multi-domain SMS spam classification
title_full ALBERT-BiLSTM cross-attention network with progressive knowledge distillation for multi-domain SMS spam classification
title_fullStr ALBERT-BiLSTM cross-attention network with progressive knowledge distillation for multi-domain SMS spam classification
title_full_unstemmed ALBERT-BiLSTM cross-attention network with progressive knowledge distillation for multi-domain SMS spam classification
title_short ALBERT-BiLSTM cross-attention network with progressive knowledge distillation for multi-domain SMS spam classification
title_sort albert bilstm cross attention network with progressive knowledge distillation for multi domain sms spam classification
topic Spam detection
BiLSTM
ALBERT
Knowledge distillation
Deep learning
Text classification
url http://www.sciencedirect.com/science/article/pii/S259012302502794X
work_keys_str_mv AT bsaparna albertbilstmcrossattentionnetworkwithprogressiveknowledgedistillationformultidomainsmsspamclassification
AT remyas albertbilstmcrossattentionnetworkwithprogressiveknowledgedistillationformultidomainsmsspamclassification
AT manujpillai albertbilstmcrossattentionnetworkwithprogressiveknowledgedistillationformultidomainsmsspamclassification
AT somularamasubbareddy albertbilstmcrossattentionnetworkwithprogressiveknowledgedistillationformultidomainsmsspamclassification
AT yongyuncho albertbilstmcrossattentionnetworkwithprogressiveknowledgedistillationformultidomainsmsspamclassification