Aug 2026· International Journal of Information Security· Vol 25· 0 citations· 35 references
Abstract
Phishing attacks via desktops, smartphones and internet of things devices are becoming increasingly sophisticated, posing critical security challenges for digital infrastructures. Defending against these attacks requires AI-based detection models that maintain high accuracy, since false positives or negatives can lead to severe breaches, while remaining lightweight enough to run on resource-constrained client devices. Split Learning (SL) meets these requirements by having clients compute only initial model layers locally and transmit intermediate activations (“smashed data”) to a server for the remaining inference, avoiding direct sharing of raw inputs. However, prior work in the image domain has shown that smashed data can leak original content, suggesting that SL may not be safe for user privacy. Therefore, it is essential to investigate whether these privacy risks also extend to language-model–based SL systems, which have fundamentally different neural network architectures, including attention mechanism. This paper introduces the Semantic Information Reconstruction Attack (SIRA), a novel framework designed to infer sensitive semantic elements directly from smashed data by leveraging the generative capabilities of large language models. In experiments on real-world phishing datasets, SIRA outperforms conventional reconstruction attacks in accurately inferring private webpage information. These findings reveal a potential privacy vulnerability in SL-based language models for security applications and motivate the development of targeted defense strategies.
The ubiquity of Application-to-Person (A2P) messaging has inadvertently created a robust vector for mobile security threats, specifically Smishing (SMS Phishing). Traditional lexical filters, such as Naive Bayes and Support Vector Machines (SVM), exhibit diminishing returns against modern obfuscated attacks due to their inability to interpret semantic context. While Transformer-based architectures like BERT have revolutionized Natural Language Processing (NLP), their computational latency renders them impractical for real-time deployment on resource-constrained mobile edge devices. This study proposes a high-efficiency detection framework utilizing DistilBERT, a distilled knowledge representation of the BERT transformer. By fine-tuning this architecture on the UCI SMS Spam Collection, we achieved a testing accuracy of 99.04% and a weighted F1-score of 0.9904. Notably, the model maintained a precision of 0.97 on the minority spam class, effectively mitigating the class imbalance problem without synthetic data augmentation. These results substantiate the viability of Knowledge Distillation as a mechanism to deploy state-of-the-art semantic security filters on edge infrastructure.
Keywords - SMS Spam, Knowledge Distillation, DistilBERT, Cybersecurity, Edge AI, NLP, Smishing, Transfer Learning
Mrinal Mrinal, Neeraj Kumar· International Journal of Cre...· 0 citations
Phishing attacks constitute one of the most pervasive cyber threats, exploiting human behaviour and system
vulnerabilities to steal sensitive information such as login credentials and financial data. Existing detection systems frequently
rely on singular modelling approaches and thus fall short in addressing the multidimensional and continuously evolving nature
of these attacks. To overcome this challenge, the present work proposes a hybrid phishing detection framework that integrates
three complementary techniques: DistilBERT (Distilled Bidirectional Encoder Representations from Transformers) for semantic
analysis of URL text, Graph Neural Networks (GNN) for modelling structural relationships among URL components, and
LightGBM (Light Gradient Boosting Machine) for efficient metadata-based feature classification. The outputs of the three
models are fused through a weighted ensemble voting mechanism, enabling the system to leverage textual, structural, and
statistical perspectives simultaneously. By replacing the computationally heavy BERT model used in prior work with the lighter
DistilBERT variant, the proposed system achieves reduced computational complexity, faster inference, and lower memory
consumption while preserving detection quality. Experimental evaluation on a large-scale Kaggle phishing URL dataset
demonstrates that the hybrid ensemble yields a weighted phishing probability of 86.9%, with individual model contributions of
0.628 (DistilBERT), 1.000 (GNN), and 0.999 (LightGBM). The system is deployed via a Flask-based web interface that enables
real-time URL classification. This multi-level approach underscores the importance of strengthening online security through
robust, efficient, and scalable detection mechanisms.
Ms. I. Shalini, Ms. G. Sujini· International Journal for Re...· 0 citations
Phishing email detection has been destabilised by Large Language Models (LLMs), which allow attackers to mass-produce fluent, contextually plausible messages. Detectors tuned on legacy corpora, notably the 2008 CEAS spam collection, consequently report benchmark accuracies that do not survive deployment. We present PhishingGAT, a detector that fuses word-level semantic features with structural ones and is hardened against adversarial perturbation. Each email is represented as an undirected graph whose nodes are unique tokens carrying 100-dimensional GloVe embeddings, and whose edges record co- occurrence inside a three-token sliding window. A two-layer Graph Attention Network (GAT) then learns per-edge importance over that graph. Robustness comes from Projected Gradient Descent (PGD) adversarial training, applied to continuous node features during optimisation only. Training uses a hybrid corpus of 70,716 emails assembled from the CEAS collection, the Enron collection, and curated LLM-generated samples. Evaluation follows two protocols. The first is strictly zero-shot: 1,186 unseen LLM-generated adversarial emails, on which the model reaches 77.40 per cent accuracy and an area under the receiver operating characteristic curve (AUC) of 0.8700. The second is domain-adapted: 401 held-out hard samples, on which accuracy rises to 90.27 per cent and AUC to 0.9717. The 12.87 percentage-point difference between the two is, to our knowledge, the first such measurement reported for graph-based phishing detection, and it quantifies a degradation that same-distribution benchmarks systematically hide. An ablation separates the contribution of attention from that of adversarial defence, and repeated runs across independent random seeds confirm that the reported figures are reproducible rather than seed artefacts.
R. Kodali, Siva Rama Krishna T Dr· International Journal of Inn...· 0 citations
Phishing remains one of the most pervasive cybersecurity threats, exploiting humanand technical vulnerabilities and targeting users through deceptive Emails,URLs, and SMS messages. Artificial Intelligence (AI) and Machine Learning (ML)techniques have been widely used to improve phishing detection accuracy. However,most existing studies have focused on specific data types, thereby limitingthe scope of their applicability, and lacking a generalized framework integratingheterogeneous data sources within phishing context. In this study, we propose ageneralized phishing detection framework that leverages classical machine learning(Random Forest and Logistic Regression) and deep learning (ConvolutionalNeural Network) to identify phishing attempts across heterogeneous textual data,such as Emails, URLs, and SMS messages. Moreover, we integrate interpretabilityinto model decisions using Explainable AI, particularly SHapley Additive exPlanations(SHAP), to enhance transparency and trustworthiness. The framework isevaluated based on both predictive performance and inference efficiency. Experimentalresults show that Random Forest achieves the highest accuracy (93%)and F1-score (85%), highlighting the efficiency of the classifier on tabular data forthe binary classification task at hand, while SHAP local and global explanationsreveal semantically relevant features influencing model decisions, where wordssuch as “admin” and “login” are identified as strong phishing indicators. These resultsdemonstrate the promise of our unified, interpretable approach in advancingadaptive and trustworthy generalized phishing detection systems.
Lea Mansour, Nour Hilal, Nadine Abbas et al.· Computer and Decision Making...· 0 citations
The security landscape is increasingly at risk from phishing attackers; individuals engaging in this type of attack successfully persuade users to divulge sensitive data such as identity information or financial resources. Existing blacklist-based solutions and other approaches have proved ineffective because the adversaries are able to adjust to current bottleneck detectors and because machine learning models are based on hand-crafted behavior patterns. Deep learning brings the inductive bias of generalization to the fore while explicitly constraining the architectures to only hold local representations and to limit their capacity to express long-range relationships in data. We present a model based on CapsNet for phishing website detection in this paper. In the framework of CapsNets, dynamic routing is used to fortify the relationship between spatial and hierarchical features in a model that is built upon these orders between features. This approach is mostly to enable the model to gain complex relationships in a URL string and HTML structure. As a result of this property, the architecture is resistant to various types of adversarial perturbations like character replacement, subdomain mimicry, and misphrasing in phishing URLs. The model is evaluated the model using datasets from PhishTank, OpenPhish, and a benign domain list sourced from the Tranco list. Experimental results show that the proposed method is superior to the state-of-the-art deep learning methods and traditional classifiers in terms of accuracy, precision, recall, and ROC-AUC and reduces the false positive rate. The results confirm that CapsNet is a stable and flexible method for phishing classification that can be used for designing and implementing secure reliable web protection systems.
Sujatha Krishna, Krishna Chandra M, Rajesh Natarajan et al.· 2026 4th International Confe...· 0 citations
Phishing is considered one of the most widespread and dynamic cyber threats, as attackers use deceitful URL structures to circumvent traditional detection tools. Despite promising performances of machine learning-based phishing detection techniques, the majority of existing models are evaluated under clean conditions, failing to account for adversarial URL evasion methods including obfuscation, encoding manipulation, homoglyph substitutions, token padding, and subdomain reordering. This weakness grossly compromises their practical performance in real-world deployments. Phishing URL detection is modelled as a lightweight binary classification problem supported by an explicit adversarial threat model. Lexical, structural, and metadata-based URL features are represented comprehensively without relying on raw URL strings or deep packet inspection (DPI), preserving user privacy and enabling broad deployment. To mitigate adversarial vulnerabilities, the Adversarial-Resilient Lightweight Random Forest (AR-LRF) model is proposed, combining controlled ensemble complexity with simulated adversarial perturbations applied during training. Extensive experiments on a large-scale imbalanced real-world dataset (91% benign, 9% phishing) consisting of 650,000 URLs were evaluated across accuracy, precision, recall, F1-score, and ROC-AUC under clean and adversarial conditions against five baseline classifiers: Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, and Naïve Bayes. The AR-LRF achieves 99.78% accuracy and ROC-AUC of 0.9999 on clean data, with significantly lower performance degradation under adversarial perturbations. Feature importance and confidence stability analyses further validate the model's resilience, positioning AR-LRF as a viable, lightweight, and deployable tool for modern cybersecurity infrastructure.
A. Chaudhuri, M. B· Scientific Reports· 0 citations