An IBGWO optimized feature selection framework for sentiment analysis-based fraudulent website detection using MLRNN Authors: Saraswathi Perumal, Anchitaalagammai Jayalakshmi Vishwanathan Published: 2026 ABSTRACT Detecting fraudulent websites is critical to ensuring network security and protecting users from online threats. Advanced machine learning algorithms are introduced to analyse patterns, anomalies, and behaviours associated with malicious online activity, which helps detect fraudulent websites. This method only detects scam sites, and there is no practical way to detect new scam websites quickly. The previous methods were not better at handling large amounts of data and could not show accurate results. In addition, fraudulent website detection poses significant threats to online security, and often struggle with high-dimensional feature spaces, feature redundancy and limited generalization to unseen attacks. To resolve these issues, this paper presents a new method for detecting fraudulent websites based on Improved Binary Grey Wolf Optimization feature subset selection based on the Multi-Layer Recognition Neural Network (MLRNN) to enhance quality features and improve overall accuracy. First, fraudulent website detection was collected from the Kaggle repository. After preprocessing, the second step is selecting the features using Attributes-based Recursive Feature Elimination (A-RFE) to select relevant features to improve model performance and minimize overfitting. Next, Subset Feature Selection involves selecting a suitable subset of features from the original dataset and a fitness function that evaluates the quality of the chosen subset using Improved Binary Grey Wolf Optimization (IBGWO). After feature selection, the data is validated using K-10-fold cross-validation, which involves repeating the training and testing procedure 10 times across different data distributions. K-10-fold cross-validation offers a sufficient approximation of the model performance. Multi-Layer Recognition Neural Network (MLRNN) is used for the sentiment analysis process due to its ability to recognize complicated data links and patterns. This model can effectively understand users’ emotion-related features to detect fraud websites efficiently. The proposed model evaluated on the web page phising detection dataset contains 11,430 URLs with 89 features. Based on the confusion matrix, the simulation results can be used to analyse precision, recall, accuracy, error rate, f-measure, and time complexity, showing higher accuracy than previous methods. Introduction Machine Learning is used to analyze patterns, anomalies, and behaviors associated with malicious online activity; machine learning plays a key role in detecting fraudulent websites. Supervised learning allows a model to be trained on labeled datasets that include examples of legitimate and fraudulent websites. These models learn to select subtle features and characteristics which differentiate fraudulent websites from genuine ones. Using machine learning for malicious website detection not only improves efficiency in detecting potential threats. Sentiment analysis and fraudulent website detection are independent applications of machine learning, and each serves different purposes. Sentiment analysis concentrates on determining the sentiment expressed in written content, such as Positive, Negative, or Neutral comments. On the other hand, scam website detection involves detecting and preventing malicious or deceptive activity on a website. Although these tasks may seem unrelated, sometimes sentiment analysis can provide additional context and functionality that can improve fraudulent website detection efforts. Integrating social media-based methods and sentiment analysis to detect fraudulent websites requires leveraging user-generated content and comments across different platforms. Here is a general approach to integrating sentiment analysis to detect scam websites on social media: Data Collection is to collect data from social media platforms, including posts, comments and discussions related to website or online activity. This data may be obtained through APIs, web scraping, or other data collection methods. Preprocessing is to clean and preprocess the text data which helps to remove noise, irrelevant details or any elements that might interfere with the analysis. Perform tasks such as tagging text, stopping word removal, and stemming. Sentiment analysis techniques are applied to evaluate the sentiment expressed in social media content. Common emotion categories include positive, negative, neutral. Tools like NLP libraries or pre-trained sentiment analysis models can be used for this task. User feedback analysis is pay attention to user feedback and comments regarding a particular website. Any negative sentiment or warnings expressed by users may indicate fraudulent activity. Combine the results of sentiment analysis with other features and concepts to build detailed ML models. Supervised ML models are trained for predicting likelihood of Internet fraud based on a combination of emotional characteristics and other relevant factors. Implement systems that continuously monitor social media channels for changes in sentiment and emerging patterns related to potential fraud. Frequently update your model with new data. By integrating sentiment analysis into fraud detection models, particularly with a focus on public network data, companies can discriminate into user opinions and concerns about specific websites. This approach increases the ability to identify and mitigate potential threats in a timely manner. Even though ML and DL methods have shown to be very useful in identifying fraudulent websites, there are various weaknesses that exist in the existing studies. The majority of the methods are based on fixed sets of features and traditional ways of feature selection that cannot be applied to high-dimensional data and quickly changing fraudulent trends. Moreover, most of the current models dwell more on structural and lexical points of the websites, as opposed to sentiment-based behavioural signals of the web sites, which are widely used by fraudulent websites to control the users. The feature selection methods that are reported in the literature based on optimisation are also characterised by premature convergence and high-level computational complexity that constrain their scalability and their application to real-life scenarios. Driven by these loopholes, this paper seeks to come up with an emotion-sensitive, optimisation-based system of detection that boosts the quality of features, improves the classification rate, and makes it resistant to the emergence of frauds sites. Despite significant progress in detecting fraudulent websites, some key challenges remain. Fraudulent websites often modify their content and presentation to evade existing detection mechanisms, thus limiting the effectiveness of traditional models. Furthermore, current methods typically rely on high-dimensional, redundant feature sets, which increase computational complexity and reduce classification performance. Furthermore, emotionally driven persuasive strategies, widely used on fraudulent websites, cannot be adequately captured by traditional feature-extraction methods. These challenges require an improved detection framework that combines sentiment analysis with improved feature selection to improve accuracy and robustness. Fig. 1 described a basic introduction to fraudulent website detection. It shows key components and process steps. Data collected initially are processed and analyzed to extract relevant features and patterns. This step deals with cleaning, transforming, and preparing data for further processing. Extract features from analyzed data, prepares data by applying preprocessing techniques and this data is fed into machine learning models. Machine Learning models’ performance is assessed by using a validation dataset.Fig. 1Basic introduction diagram for fraudulent website detection. Basic introduction diagram for fraudulent website detection. A promising approach for detecting website fraud is the use of Multi-Layer Recurrent Neural Networks (MLRNNs). The proposed approach combines the advantages of a multi-layer architecture for identifying complex relationships in data and RNNs, which are specifically designed to capture sequence-dependent dependencies. The proposed feature sets are examined, and several classification techniques are used. Based on the implementation results, each of the proposed attributes are useful for phishing detection. Contribution of this work The main contributions of this work are summarized as follows:A novel sentiment analysis–based fraudulent website detection framework is proposed that integrates textual sentiment cues with structural website features.An IBGWO algorithm is employed for feature subset selection to effectively reduce feature redundancy and computational complexity.An A-RFE is combined with IBGWO to enhance the quality of selected features prior to classification.The optimized feature set is classified using a MLRNN, resulting in improved detection accuracy validated through 10-fold cross-validation. A novel sentiment analysis–based fraudulent website detection framework is proposed that integrates textual sentiment cues with structural website features. A novel sentiment analysis–based fraudulent website detection framework is proposed that integrates textual sentiment cues with structural website features. An IBGWO algorithm is employed for feature subset selection to effectively reduce feature redundancy and computational complexity. An IBGWO algorithm is employed for feature subset selection to effectively reduce feature redundancy and computational complexity. An A-RFE is combined with IBGWO to enhance the quality of selected features prior to classification. An A-RFE is combined with IBGWO to enhance the quality of selected features prior to classification. The optimized feature set is classified using a MLRNN, resulting in improved detection accuracy validated through 10-fold cross-validation. The optimized feature set is classified using a MLRNN, resulting in improved detection accuracy validated through 10-fold cross-validation. Unlike conventional fraudulent website detection pipelines, the proposed approach introduces a two-stage feature selection strategy that combines A-RFE and IBGWO before classification. This integration enables effective handling of high-dimensional, heterogeneous features while preserving sentiment-related information, which the MLRNN classifier leverages to enhance fraudulent website detection performance. In this study, we present a novel integrated approach of sentiment analysis, IBGWO-based feature optimization, and MLRNN classification for fraudulent website detection. The proposed framework explicitly addresses feature redundancy and high dimensionality, which are often overlooked by existing approaches, improving detection robustness and computational efficiency. This study will be trying to answer the following research questions based on the identified research gaps:(i)What are the best methods to integrate the sentiment-related features in detecting fraudulent websites such that the classification performance can be enhanced?(ii)Does Attributes-based Recursive Feature Elimination with Improved Binary Grey Wolf Optimization do better at selecting a more optimum feature subset than traditional selection methods?(iii)Does a Multi-Layer Recognition Neural Network that has been trained on optimised subsets of features yield a higher detection accuracy and generalisation capacity?(iv)What is the strength of the proposed model when tested with different data distributions by means of k-fold cross-validation? (i)What are the best methods to integrate the sentiment-related features in detecting fraudulent websites such that the classification performance can be enhanced? What are the best methods to integrate the sentiment-related features in detecting fraudulent websites such that the classification performance can be enhanced? (ii)Does Attributes-based Recursive Feature Elimination with Improved Binary Grey Wolf Optimization do better at selecting a more optimum feature subset than traditional selection methods? Does Attributes-based Recursive Feature Elimination with Improved Binary Grey Wolf Optimization do better at selecting a more optimum feature subset than traditional selection methods? (iii)Does a Multi-Layer Recognition Neural Network that has been trained on optimised subsets of features yield a higher detection accuracy and generalisation capacity? Does a Multi-Layer Recognition Neural Network that has been trained on optimised subsets of features yield a higher detection accuracy and generalisation capacity? (iv)What is the strength of the proposed model when tested with different data distributions by means of k-fold cross-validation? What is the strength of the proposed model when tested with different data distributions by means of k-fold cross-validation? Related work Related work can analyze the methods that various researchers have suggested utilizing Data Mining, ML and DL techniques to detect malicious websites. Success of these techniques depends on the quality and composition of relevant aspects of the website fraudulent detection. GandotraEet al.1 defined as, one of the main concerns for security researchers nowadays are the rise in phishing assaults. Traditional phishing website identification tools use signature-based methods and are unable to detect newly created phishing web pages. F. A. Ghaleb et al.2 defined as, most credit card fraud occurs online through data breaches, phishing and fraud to obtain payment credentials. To address the problem of credit card theft in internet purchases, numerous solutions have been offered. However, higher-order inequalities are a major challenge for existing solutions in building effective detection models. S. Makki et al.3 defined as, the theft of credit cards is a severe offense. Both people and financial institutions are significantly impacted. Financial firms must consequently prioritize the detection and prevention of fraud. Fraud detection and prevention is an expensive, labor-intensive, and time-consuming process. D. Cheng et al.4 described card fraud as a serious issue that costs both cardholders and the institutions that furnish the cards a lot of money. To identify fraud in transaction records, modern methods rely on machine learning. J. Chen et al.5 defined as, Provide a healthcare industry-relevant Automated Medical Fraud Detection (AMFD) framework. To solve this issue, variety of medical fraud labels is insufficient to train a good AMFD model. However, these methods rely heavily on manual feature engineering and often struggle with high-dimensional and redundant feature spaces, limiting their scalability and generalization to unseen phishing patterns. M. Grossi et al.6 defined as, IBM Safer Payments and the Kiskit software stack were utilized to apply QSVM algorithms to classification problems in the financial payments industry for the first time, using IBM Quantum processors. Using real card payment data, a thorough investigation is carried out to evaluate the complementing impacts of cutting-edge quantum machine learning algorithms to conventional techniques. W. Ning et al.7 defined as, the classic Ad boost algorithm’s robustness has been improved. The classic AdaBoost algorithm’s low robustness is due to the fact that just the error rate is determined using the basic classifier. Nevertheless, these approaches typically demand large labeled datasets, incur high computational costs, and are sensitive to noisy or irrelevant features, which can degrade performance in real-world environments. Y. Yoo et al.8 described as, Models for detecting health insurance fraud use two techniques to describe graph information: models of Graph Neural Networks (GNNs) and traditional machine learning models that make use of graph centrality metrics. . Dutta AK et al.9 defined as, Phishing attacks are common in nature in which victims are duped into viewing dangerous content and revealing personal information. Most legitimate websites and phishing pages are similar in terms of their user experience and uniform resource location (URL). Purbay M et al.10 defined as, identify phishing URLs by using deep learning and machine learning techniques to assess multiple URL components.We examine supervised learning approaches for detecting phishing URLs using lexical characteristics, WHOIS characteristics, PageRank, Traffic Rank profiles, and Page Importance attributes. Jain A.K et al.11 defined as, PHISH-SAFE, an anti-phishing solution that uses machine learning and the URL function. In order to evaluate the effectiveness of our suggested method, we took 14 factors out of URLs to identify phishing websites. However, these models are primarily designed for sentiment polarity classification and are not optimized for web-content-oriented phishing detection, limiting their effectiveness when applied directly to fraudulent website identification. According to Rao RS et al.12, phishing is the act of utilizing false websites that replicate legitimate ones in order to get personal data such as social security numbers, credit card information, usernames, and passwords. To combat phishing attacks, the latest methods utilize search engine-based approaches with assured detection accuracy. AlEroud A et al.13 defined as, when creating phishing detection systems, the web address URL is typically utilized. Unusual patterns in URLs are frequently recognized by machine learning algorithms as possible signs of phishing.By developing samples that work around the classification algorithm, an attacker might, nevertheless, have sufficient expertise and drive to get around the URL classification procedure. Sahingoz OK et al.14 defined as, very challenging to tell if a website is phishing or genuine, as web pages are primarily semantic-based attack structures that exploit computer user vulnerabilities. Kazeem et al.15 defined as, use to increase fraud identification and prevention in financial transactions, real-time monitoring systems and machine learning algorithms are being used. Vanjari et al.16 defined as, Rank Fraud is the practice of engaging in fraudulent or deceptive practices with the intent of artificially increasing an app’s ranking on a popular list. In fact, ranking fraud is becoming increasingly common by app developers. These activities include exaggerating the sales of apps and uploading fake app reviews. Table 1 describes the traditional methods for Fraudulent website detection with their limitations.Table 1Traditional methods for fraudulent website detection.Author/YearProposed methodLimitations17Visual analytics approach for Collusive fraud detectionLack of labeled data in systematic clinical visits18CoDetect fraud detectionIn the financial sector, fraud is hard to spot, and the characteristics of fraud are difficult to identify.19Process-based fraud (PBF) using Process miningPBF innovation indicators now available do not effectively answer the needs of varied conceptual viewpoints of business processes.20Convolutional neural network (CNN)Due to the low accuracy, it should be used to reduce the losses due to fraud.21ReMEMBeRFraud detection has a low adaptability and a high false positive rate22Dolos , a fraud de-anonymization systemCannot identity human brains that publish fraudulent search rankings on peer-reviewed websites.23Logistic Regression (LR), Support vector machines (SVM)It is characterized by a large difference in monitoring frequency between fraudulent and non-fraudulent transactions, and contains irrelevant, irrelevant, and highly correlated data that negatively affect forecasting performance.24Generative adversarial networks(GAN)Unbalanced data sets have become a serious issue.25CS-OCANWhen handling complicated and unpredictable situations, they usually prove to be ineffectual.26LGM-GNNNeighbor information not valid or aggregating to detecting the fraud websites. Table 1Traditional methods for fraudulent website detection.Author/YearProposed methodLimitations17Visual analytics approach for Collusive fraud detectionLack of labeled data in systematic clinical visits18CoDetect fraud detectionIn the financial sector, fraud is hard to spot, and the characteristics of fraud are difficult to identify.19Process-based fraud (PBF) using Process miningPBF innovation indicators now available do not effectively answer the needs of varied conceptual viewpoints of business processes.20Convolutional neural network (CNN)Due to the low accuracy, it should be used to reduce the losses due to fraud.21ReMEMBeRFraud detection has a low adaptability and a high false positive rate22Dolos , a fraud de-anonymization systemCannot identity human brains that publish fraudulent search rankings on peer-reviewed websites.23Logistic Regression (LR), Support vector machines (SVM)It is characterized by a large difference in monitoring frequency between fraudulent and non-fraudulent transactions, and contains irrelevant, irrelevant, and highly correlated data that negatively affect forecasting performance.24Generative adversarial networks(GAN)Unbalanced data sets have become a serious issue.25CS-OCANWhen handling complicated and unpredictable situations, they usually prove to be ineffectual.26LGM-GNNNeighbor information not valid or aggregating to detecting the fraud websites. Traditional methods for fraudulent website detection. Adoma AF et al.27 defined as, Assess the effectiveness of XLNet, RoBERTa, DistilBERT, and BERT pre-trained transformer models in text sentiment detection. Analyzing the output of each candidate model against that of other candidate models. Machova et al.28 defined as, Machine learning techniques are used to develop detection models to differentiate between troll commenters and regular commenters, and sentiment analysis techniques are also used to determine the general tone of troll comments. Sentiment analysis can also be done with machine learning or dictionary-based techniques. Meenu Mathew et al.29 defined as, Text sentiment detection is a classification issue in which different sentiments are assigned to diverse text inputs. This is a difficult undertaking due to the heterogeneity and uncertainty. Existing machine learning approaches can recognize emotions, but they are incapable of handling very long durations. Dashtipour, K et al.30 defined as, is based on symbolic linkages, the framework permits emotions to flow from words to concepts. If the mode is not triggered, the framework switches to the corresponding sub-symbol and classifies using a Deep Neural Network (DNN). m Al Amrani Y et al.31 describe sentiment analysis from product reviews using a unique text search method. Depending on some aspect of the time-based query, these reviews can be categorized as positive or negative sentiments. Alqaryouti O et al.32 defined to overcome sentiment analysis issues and create abstract findings, this approach employs language processing tools, rules, and dictionaries. According to the results, taking into account hidden features enhances feature extraction accuracy greatly. Furthermore, the combined classification model performed 5% better than the combination of the dictionary-based basis and additional criteria. Meena et al.33 presents an FSTL-SA is a few-shot transfer learning architecture that classifies sentiment polarity using nonverbal facial expression and that generalises well to the CK+ dataset as well as FER2013 dataset. Nevertheless, this methodology is concerned with facial sentiment analysis but not text-based sentiment features, feature optimisation and fraudulent websites. On the contrary, the suggested approach combines sentiment analysis of web content and the IBGWO-based feature selection to enhance the detection of fraud. Choudhary et al.34 introduced a deep-learning-based solution to the problem of image dehazing, which notes that neural networks can be trained to learn to distinguish discriminative features of noisy visual inputs. Even though the field of application is different, this work supports the usefulness of deep learning structures in determining the meaningful patterns in complex data. The proposed study is driven by these abilities and uses a multi-layer neural network to predict pattern of sentiment in fraudulent web contents. Meena et al.35 suggested the framework of a Point of Interest recommendation that incorporates the sentiment prediction based on a BiLSTM model that is trained on an election-based dataset. The strategy shows that sentiment-based learning is effective in individualized recommendations. Nevertheless, the model is concerned with the accuracy of recommendations and does not consider feature selection optimization, fraud detection, and the high-dimensional web content analysis. The proposed work, in its turn, uses sentiment analysis in detecting fraudulent sites and uses feature subset selection based on optimization in order to increase the robustness of classification.Ramkumar et al. [36] introduced the Subset Clustering-based FS using MLPNN (SCFS-MLPNN) approach for cancer risk detection. The technique applied IMDIR as a preprocessor, SDPSR as a disease pattern generator, and the ICFSS technique for feature clustering and selection based on marginal rates. The chosen features were classified using an MLPNN, resulting in improved prediction accuracy for cancer risk and facilitating early treatment. Kalpana et al. [37] proposed an IoT-based smart farming system using cloud computing and machine learning for precision agriculture. The system runs on a Raspberry Pi, which gathers sensor data and crop images. Image preprocessing was performed using Wiener filtering followed by region-based segmentation. Ensemble-based classification was used to classify plant diseases, predict crop yield, and identify insect damage for effective real-time farm monitoring. Assegie [38] introduced a K-Nearest Neighbour (KNN)-based phishing detection model using PhishTank data. The approach categorises URLs as phishing or legitimate based on similarity measures. The findings revealed that KNN outperforms other methods in correctly classifying URLs as phishing attacks. Problem statement Many of the existing methods 11, 12, and 13 are hand-crafted, such as the number of dots, the presence of special "@," "#," and "-" symbols, the length of the URL, the brand name in the URL, and high-level extracting URL-based features using manual URL feature extraction still presents significant hurdles, as it requires man-time and additional maintenance work costExperimental results indicate that the technology relies on machine learning and may struggle with large volumes of data.Detection method has been used to detect (not classify) malicious and valid web pages. The results show that there is a greater true positive rate than false positive rateLanguage issues – Items that users do not understand may be wrongly recommended (for example, if the language of the label is different from the language of the content). It is considered very useful for identifying user interests, but it isn’t easy to achieve high efficiency Many of the existing methods 11, 12, and 13 are hand-crafted, such as the number of dots, the presence of special "@," "#," and "-" symbols, the length of the URL, the brand name in the URL, and high-level extracting URL-based features using manual URL feature extraction still presents significant hurdles, as it requires man-time and additional maintenance work cost Many of the existing methods 11, 12, and 13 are hand-crafted, such as the number of dots, the presence of special "@," "#," and "-" symbols, the length of the URL, the brand name in the URL, and high-level extracting URL-based features using manual URL feature extraction still presents significant hurdles, as it requires man-time and additional maintenance work cost Experimental results indicate that the technology relies on machine learning and may struggle with large volumes of data. Experimental results indicate that the technology relies on machine learning and may struggle with large volumes of data. Detection method has been used to detect (not classify) malicious and valid web pages. The results show that there is a greater true positive rate than false positive rate Detection method has been used to detect (not classify) malicious and valid web pages. The results show that there is a greater true positive rate than false positive rate Language issues – Items that users do not understand may be wrongly recommended (for example, if the language of the label is different from the language of the content). It is considered very useful for identifying user interests, but it isn’t easy to achieve high efficiency Language issues – Items that users do not understand may be wrongly recommended (for example, if the language of the label is different from the language of the content). It is considered very useful for identifying user interests, but it isn’t easy to achieve high efficiency Materials and method Sentiment-based fraud website detection increases the precision and effectiveness of identifying fraudulent websites by utilizing Multi-Layer Recognition Neural Network (MLRNN). RNN algorithms analyze emotion expressed in user reviews, comments, and other website-related textual content to find subtle indicators of fraudulent activity. The goal of using IBGWO for subset feature selection is to quickly find the appropriate features from a given set which will enhance machine learning model performance and reducing the computational complexity. Sentiment-based MLRNN approach enables the development of highly efficient and accurate detection systems. Fig. 2 describes a proposed diagram for Detecting fraud websites based on user sentiment analysis. Initially, we collected the dataset from the Kaggle repository based on user sentiment analysis and Fraud website detection. Start the preprocessing section. It includes two steps: text preprocessing and Data scaling, used to scale the data values and process the text or numerical values. The second step is featuring selection based on A-RFE for finding the relation features. The third one is subset features selection, which is mainly to select the sub-features based on the IBGWO algorithm. By using K-10 fold validation, Validate the test and training dataset before classification. In classification, MLRNN algorithm can be helpful for detecting fraudulent websites using sentiments.Fig. 2Proposed flow diagram. Proposed flow diagram. Dataset description This research uses the Web Page Phishing Detection dataset is gathered from Kaggle repository. The available link is https://www.kaggle.com/datasets/shashwatwork/web-page-phishing-detection-dataset. This dataset file size is 3.66MB. The supplied dataset contains 11,430 URLs, from which 87 features were extracted. Features were extracted into three categories: 56 from the URL’s structure and syntax, 24 from the page’s content, and seven from external services. This dataset is balanced, consisting of 50% phishing URLs and 50% legitimate URLs. Despite its suitability as a benchmark, the Web Page Phishing Detection dataset presents several challenges. The dataset contains a high-dimensional feature space with 87 heterogeneous attributes derived from URL structure, webpage content, and external services, which increases redundancy and computational complexity. Features obtained from external services may introduce latency and dependency issues in real-time applications. Additionally, although the dataset is balanced, real-world phishing scenarios are often highly imbalanced, which may affect model generalization. These challenges motivate the need for effective feature selection and optimization techniques, as addressed in the proposed IBGWO-based framework. Preprocessing stage One of the most important steps in developing a model for detecting fraudulent websites is pre-processing. In order to train machine learning models, raw data must be cleaned and formatted. Find missing values, duplicate values and inconsistencies in data sets. Eliminate extraneous or redundant activity that does not contribute to the innovation process. Standardize or measure numerical characteristics that fit within a fixed range. This ensures that the large-scale features do not dominate the learning process. Text Pre-processing Text preprocessing is the process of executing a series of mathematical operations and transformations to change raw text input into a format that is acceptable for machine learning models. Prepare and organize text data. For examples, Word segmentation, stop word elimination, word formation, and lemmatization. Text preprocessing steps into mathematical equations are, Tokenization Tokenization is a process of breaking up the original text into discrete words or symbols. This operation is performed by the tokenize function.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Tokens=Tokenizer(Raw \,\,Text)$$\end{document} Removing stop words Stop words can be removed by using the function RemoveStopWords from the tags. Common words such as “the” and “and”.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Filtered\,\, Tokens=RemoveStopWords(Tokens)$$\end{document} stemming/lemmatization Words can be reduced to their simplest form by Stemming or Lemmatization, while features and stemming/lemmatization can be done for filtered tokens.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Processed\,\ Tokens=Stemming/Lemmatization(Filtered \,\,Tokens)$$\end{document} Vectorization (word embedding) Single word embeddings convert processed tokens into dense vectors. The Word Embedding function creates the embedding array.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Embedding\,\, Matrix=WordEmbedding(Processed\,\, Tokens)$$\end{document} Text vectorization (TF-IDF): Weights are assigned to tags by the TF-IDF vectorization according to their significance within the document. TF-IDF vectorization function produces the TF-IDF matrix.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TF-IDF Matrix=TFIDFVectorization(Processed\,\, Tokens)$$\end{document} These procedures can be implemented in a variety of ways, and many libraries and tools (like TensorFlow and scikit-Learn) have the capability to perform these changes in a machine-readable format efficiently. Once the preprocessing of the text is done, the clean textual data are converted into numerical forms through word embedding algorithms. Semantic and contextual relationship between words in web-based information is captured using pre-trained embedding models to represent sentiment. The aggregation of word-level embeddings represents each site as a fixed-dimensional embedding vector. These vectors are then added to other features extracted and sent to the feature selection phase where A-RFE and IBGWO will find the most discernible embedding dimensions to detect a fraudulent internet site. Data scaling Data scaling is a common preprocessing step that transforms the values of numerical features into equivalent scaling ratios. One data scaling technique that produces a mean of 0 and a standard deviation of 1 from the values of a numerical variable is called Z-score scaling, often known as standardization. This is especially helpful when utilising machine learning techniques, as these algorithms have uniformly sized, regularly distributed features. Z-score extended formula:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Z}_{s}=\frac{a-\mu }{\sigma }$$\end{document} In Eq. (1), where\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Z}_{s}-Standarized value of the data point$$\end{document}a-original valuesμ -mean valueσ -standard diviation values Calculate the Mean (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\mu}}$$\end{document}) and standard deviation (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\sigma}}$$\end{document})2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu =\frac{1}{x}\sum_{a=1}^{x}{U}_{a }$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma =\sqrt{\frac{1}{x}\sum_{a=1}^{x}{(U}_{a}-\mu {)}^{2}}$$\end{document} In Eq. (2) and (3), where. x- number of data points Apply Z-score scaling \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Z}_{s}=\frac{a-\mu }{\sigma }$$\end{document} . The variable distribution that results from this scaling has a 0 mean and a 1 standard deviation. When comparing variables at various scales, this normalization is helpful and can aid in the convergence of some machine learning algorithms during training.In Python, the Z-score scale can be applied using libraries like scikit-learn. Feature selection using attributes-based recursive feature elimination (A-RFE) Attribute-based feature selection focuses on evaluating the importance of individual attributes that are associated with fraud websites. Use A-RFE such as correlation analysis, mutual information, and statistical tests to measure relationship between each characteristic and resultant variable. In this case, it classifies the website as either scam or legitimate. Attribute-based Recursive Feature Elimination (A-RFE) is a technique which systematically removes attributes from a data set based on their importance in predicting a target variable. The attribute set of features is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F = \left\{{F}_{1},{F}_{2}, ..., {F}_{x}, ..., {F}_{n}\right\}, {F}_{n}$$\end{document} represents the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${n}^{th}$$\end{document} feature attributes, where N is the number of attributes. Recursive feature R \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F = \{{R}_{1},{R}_{2}, ..., {R}_{x}, ..., {R}_{m}\},$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{m}$$\end{document} represents the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${n}^{th}$$\end{document} category, m is the number of categories. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{n}$$\end{document} The sample size of a specific category \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$, {F}_{n}$$\end{document}, CFnis the number of feature values. The attribute values distribution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{n}$$\end{document} for the feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{m}$$\end{document} can be expressed as:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${FR}_{nm}=\left\{{F}_{x1}^{n},{F}_{x2}^{n},{F}_{x3}^{n}\dots .{FE}_{x4}^{n}\right\}$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Fi}_{n}^{rf}=|{F}_{x1}^{nr}-{F}_{x1}^{nf}{|}_{2}=\sum_{x=1}^{{N}_{f}}({R}_{xy}^{r}-{R}_{xy}^{s}{)}^{2}$$\end{document} In Eq. no (4) denotes the maximum and minimum values of features set is that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{xy}^{r}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{xy}^{s}$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{x1}^{nr}-{F}_{x1}^{nf}$$\end{document} Feature relatable values of overall feature set mentioned in eq. no (5).6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{xy}=\frac{\sum_{x-1}^{n}{R}_{xy}^{r}}{n-1}$$\end{document} When n is an equal value to the x, the values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{xy}^{r} is$$\end{document} expressed in the eq. no (6) and defined the cumulative average of recursion degree of all categories set continuous under all feature qualities the x*y dimension matrix of R which is mentioned in the eq. no (7):7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R=\left(\left[\begin{array}{cc}{R}_{11}& {R}_{12}\dots {R}_{1n}\\ {\begin{array}{c}{R}_{12}\\ \vdots \\ R\end{array}}_{n1}& {\begin{array}{c}{R}_{22}\\ \vdots \\ R\end{array}}_{n2}{R}_{xy}{\begin{array}{c}{R}_{2n}\\ \vdots \\ R\end{array}}_{nn}\end{array}\right]\right)$$\end{document} Feature selection is implemented through a discriminant matrix Q. First, we select the most unique feature attributes. To assure the balance of feature qualities, develop a base feature set for each category. Calculate the correlation between each feature attribute in phi and feature in Relational feature base to complete the base feature set Relational feature base expansion. Then, use the average correlation between the features to choose the appropriate feature attributes.8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{xy}=\frac{coR({R}_{x},{R}_{y})}{\sqrt{var\left({R}_{x}\right) var({R}_{y}) }}$$\end{document} In eq. (8), where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$coR({R}_{x},{R}_{y})$$\end{document} is the correlational of features \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{x} and {R}_{y}$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$var({R}_{y})$$\end{document} variance of the features9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{xy}=\frac{1}{N}\sum_{y=1}^{{R}_{x}}\frac{coR\left({R}_{x},{R}_{y}\right)}{\sqrt{var\left({R}_{x}\right) var\left({R}_{y}\right)}}$$\end{document} In eq. (9), where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{x}$$\end{document} is the amount of attribute features in the base feature set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{xvar}$$\end{document}. The feature attributes with the least correlation are chosen based on the value of R, and eventually a feature attribute is set.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{R}\mathrm{F}\mathrm{E}(\mathrm{X},\mathrm{y})=\mathrm{a}\mathrm{r}\mathrm{g}\mathrm{m}\mathrm{i}\mathrm{n}\mathrm{X}^{\prime} \mathrm{P}\mathrm{e}\mathrm{r}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{m}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{M}\mathrm{e}\mathrm{t}\mathrm{r}\mathrm{i}\mathrm{c}(\mathrm{M}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{l}(\mathrm{X}^{\prime},\mathrm{y}))$$\end{document} Set of Relational features providing high accuracy\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\to F}^{R}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$If {E}_{n}^{cor}>{E}_{n-1}^{cor}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}^{R}=R$$\end{document} End. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E}_{n}^{cor}$$\end{document} –Eliminating the correlation variables, R- Recursive values, the final feature set (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E}_{n}^{cor}$$\end{document}) is defined by the best coefficient local model for predicting out-of-bag models. After obtaining the final feature set, a K-10-fold cross-validation (CV) was performed on RFR using the same settings used to train the feature removal process. Improved binary grey wolf optimization feature subset selection Improved Binary Grey Wolf Optimisation (IBGWO) is utilised in feature selection to determine the feature subset that maximises classification accuracy while minimising the number of selected features. As a binary string, represent a subset of each conceivable characteristic. Each bit determines whether the feature is included (1) or excluded (0). Set Population: Initialize the number of binary strings representing the different feature subsets. Start: Start position of binary string (wolf) in search space. Match Evaluation: Evaluate the match of each binary string using a defined match function. Rank: Rank binary strings based on matching. Upgrade Alpha, Beta, and Delta: Alpha, Beta, and Delta upgrades wolves’ status based on their health. Search for new solutions: adjust the wolf state to create new candidate solutions. Boundary Handling: Ensuring that binary strings fall within a defined search space. After convergence or a specified number of iterations, a binary string (feature subset) corresponding to the alpha wolf is extracted, representing the best solution found.Machine learning model can be trained by using features subsetchosen from the training dataset. In the IBGWO mathematical model, the optimal solution is known as alpha, beta and delta are the second and third best answers, respectively. The remaining possible solutions are rated as omega (). Predation is influenced by, and, and follows these three values.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{A}{\to }\left(X+1\right)=\underset{{A}_{n}}{\to }\left(X\right)+\underset{C}{\to }.\underset{D}{\to }$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{D}{\to }=|\underset{C}{\to }.\underset{{A}_{n}\left(X\right)}{\to }-\underset{C(X)}{\to }|$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{D}{\to }=2x.\underset{D1}{\to }-x,\underset{C}{\to }=2\underset{{C}_{2}}{\to }$$\end{document} During an iteration, it drops linearly from 2 to 0. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C}_{2}$$\end{document} are random vectors [0, 1]. Usually, the hunt is led by an alpha. Betas and deltas may occasionally engage in hunting. (The best candidate solution), (the second-best candidate solution), and (the third best candidate solution) are meant to have strong knowledge of prospective locations in order to mathematically simulate the hunting behaviour of grey wolves. With the top three best candidate solutions received so far, Omega and other search agents should adjust their positions based on the best search agent position. Therefore, update of the wolf condition is shown in the following equation.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{\mathrm{A}}{\to }\left(\mathrm{X}+1\right)=\frac{\underset{\mathrm{A}1+{\mathrm{A}}_{2+{\mathrm{A}}_{3}}}{\to }}{3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{{A}_{1}}{\to }=\left|\underset{{A}_{\propto }}{\to }+\underset{{A}_{\beta }}{\to }+\underset{{A}_{\sigma }}{\to }\right|$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{{A}_{1}}{\to }=\left|\underset{{A}_{\alpha }}{\to }-\underset{{A}_{1}}{\to }.\underset{{C}_{\alpha }}{\to }\right|$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{{A}_{2}}{\to }=\left|\underset{{A}_{\beta }}{\to }-\underset{{A}_{2}}{\to }.\underset{{C}_{\beta }}{\to }\right|$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{{A}_{3}}{\to }=\left|\underset{{A}_{\gamma }}{\to }-\underset{{A}_{1}}{\to }.\underset{{C}_{\gamma }}{\to }\right|$$\end{document}Where −\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{\mathrm{A}1+{\mathrm{A}}_{2+{\mathrm{A}}_{3}}}{\to }$$\end{document} are for any given iteration, the first three best answers in the group t, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{\mathrm{C}1+{C}_{2+{\mathrm{C}}_{3}}}{\to }$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{{C}_{\alpha }}{\to }=\left|\underset{{C}_{1}}{\to }.\underset{{A}_{\propto }}{\to }-\underset{D}{\to }\right|$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{{C}_{\beta }}{\to }=\left|\underset{{C}_{1}}{\to }.\underset{{A}_{\beta }}{\to }-\underset{D}{\to }\right|$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{{C}_{\gamma }}{\to }=\left|\underset{{C}_{1}}{\to }.\underset{{A}_{\gamma }}{\to }-\underset{D}{\to }\right|$$\end{document} According to the equation, the factor ais linearly updated between 2 and 0 in each iteration. For each \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w}_{x}\in feature do$$\end{document} Updated the current wolf position. Update A, D and C Evaluating the position of individual features. Update the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha ,\beta ,and\gamma$$\end{document} End. In bGWO, the solvation pool is binary at any given time. The solutions lie at the corners of the hypercube. Updated wolf locations in accordance with GWO principles while maintaining binary constraints.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{n}^{A+1}=CrossOver\left({A}_{1},{A}_{2},{A}_{3}\right)$$\end{document} In eq. no (10), where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CrossOver({A}_{1},{A}_{2},{A}_{3})$$\end{document} is the exact intersection between a,b,c and the solutions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${A}_{1},{A}_{2},{A}_{3}$$\end{document} are binary vectors that represent the effects of wolf alpha, beta, and delta moving toward gray wolves.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_{1}^{G} = \left\{ {\begin{array}{*{20}c} {1~~~~~~~~~~~~~~~~if\left( {a_{\alpha }^{c} + bstep_{\alpha }^{c} } \right) \ge 1} \\ {0~~~~~~~~~~~~~~~~~~~~~~~~~Otherwise} \\ \end{array} ~} \right.$$\end{document} In eq. no (11), where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${a}_{\alpha }^{c}$$\end{document}- The vector representing the alpha wolf’s position, D-Dimension, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${bstep}_{\alpha }^{c}$$\end{document} It is a binary dimension d step that can be calculated as:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$bstep_{\alpha }^{c} = \left\{ {\begin{array}{*{20}c} {1~~~~~~~~~~~~~~~~if\left( {a_{\alpha }^{c} + Cstep_{\alpha }^{c} } \right) \ge rand} \\ {0~~~~~~~~~~~~~~~~~~~~~~~~~Otherwise} \\ \end{array} } \right.$$\end{document} In eq. no (12), rand denotes uniformly distributed random number \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\in \left[\mathrm{0,1}\right], {cstep}_{\alpha }^{c}$$\end{document} is the continuous values step size for dimension (D),13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Cstep}_{\alpha }^{c}=\frac{1}{1+{e}^{-10}\left({A}_{1}^{c}{D}_{1}^{c}-0.5\right)}$$\end{document} In eq. no (13), where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${A}_{1}^{c}{D}_{1}^{c}$$\end{document} Is calculating the dimensions (D)14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$bstep_{\alpha }^{c} = \left\{ {\begin{array}{*{20}c} {1~~~~~~~~~~~~~~~~ifCstep_{\alpha }^{c} \ge rand} \\ {0~~~~~~~~~~~~~~~~~~~~~~~~~~~Otherwise} \\ \end{array} } \right.$$\end{document} In eq. no (14), an integer selected at random from a uniform distribution is called a rand. ε [0, 1], and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Cstep}_{\alpha }^{c}$$\end{document} is a d-dimensional continuous-valued step that can be determined using the sigmoid function presented in Equations.15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Cstep}_{\alpha }^{c}=\frac{1}{1+{e}^{-10}({A}_{1}^{c}{D}_{1}^{c}-0.5)}$$\end{document}16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a_{3}^{c} = \left\{ {\begin{array}{*{20}c} {1~~~~~~~~~~~~~~~~if\left( {x_{\infty }^{c} + bstep_{\delta }^{c} } \right) \ge 1} \\ {0~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~Otherwise} \\ \end{array} } \right.$$\end{document} In eq. no (15), where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${A}_{1}^{c}{D}_{1}^{c}$$\end{document} is the delta Wolff state vector in d dimension and in eq. (16), bstepd δ is the binary step size in d dimension, calculated as in the eq. no (17):17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${a}_{d}=\left\{\begin{array}{c}{x}_{d} if rand<1/2\\ {y}_{d}\frac{1}{2}\le rand<2/3\\ {z}_{d} othewise\end{array}\right.$$\end{document} The updated gray wolf state vector should be binary which is calculated by using the eq. no (18).18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a_{x}^{{c + 1}} = \left\{ {\begin{array}{*{20}c} {1~~~~~~~~~~~~~~~~if~simoid~\left( {\frac{{x_{1} + x_{2} }}{2}} \right) \ge 1} \\ {0~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~Otherwise} \\ \end{array} } \right.$$\end{document} With the fewest attributes specified, the ideal feature combination provides the best classification performance. A fitness function is used in binary grey wolf optimization to estimate the individual grey wolf states.19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Fitness={\alpha}_{\gamma R}\left(X\right)+\beta \frac{|D-R|}{|D|}$$\end{document} In eq. no (9), where γR(D) is the conditional feature set of D’s classification quality; R defines the length of the subset feature that was chosen; C defines the number of features; and two parameters are α and β subset attributes. By using IBGWO for feature subset selection in malicious website detection, the aim is to identifysmall features set that maximizes the distinction between legitimate and malicious websites. The algorithm’s exploration and exploitation capabilities help effectively navigate the binary search space. The extracted feature points from each image are denoted in fig.3. Here, n is the feature vector dimension. After extracting feature points, the IBGWO algorithm in Fig.3 is used to choose the finest features. The algorithm is initialized with a set of random solutions with binary associated values.Fig. 3IBGWO for subset feature selection. IBGWO for subset feature selection. K-10 fold cross validation To assess machine learning models on limited data samples, resampling is done by using cross validation technique. By using K-10-fold cross validation, the original dataset can be divided into K-folds (or subsets) of equal size and this K technique is helpful to train the model. Every time, practice is done on the K–10 fold, and testing and verification are done on the remaining folds. There are K iterations of this method, with each K–10 fold serving as validation data just once. One can then create a single estimate of the model performance by averaging the K-fold results. Use the entire data set for training and validation to maximize data usage.Averaging the results across multiple folds can provide a more reliable estimate of model performance.Helps to identify potential issues with over-fitting or under-fitting. Averaging the results across multiple folds can provide a more reliable estimate of model performance. Averaging the results across multiple folds can provide a more reliable estimate of model performance. Helps to identify potential issues with over-fitting or under-fitting. Helps to identify potential issues with over-fitting or under-fitting. Fig. 4 shows input and output parameters, their units, minimum and maximum values, and means. All usage data were obtained using a 10-fold cross-validation method.Fig. 4K-10 fold validation. K-10 fold validation. Multi-layer recognition neural networks (MLRNN) MLRNN design carefully integrates the features like Uniform Resource Locator structure, Domain Age, IP address reputation, Secure Socket Layer certificate details and the content characteristics to detect fraudulent websites. Neural network architectures usually consist of input layers representing relevant features, hidden layers for nonlinear transformations, and output layers for binary classification (fraudulent or legitimate). Through an iterative training process using optimization techniques such as smooth gradient descent, MLRNN learns to identify subtle patterns that may indicate fraudulent activity. Assessed by metrics such as accuracy, sensitivity, specificity using Confusion Matrix, for MLRNN demonstrated its effectiveness in distinguishing fraudulent and legitimate websites. MLRNN is well suited for processing air indexing dataset due to its high-performance ability. The symbol of A, B is the Feature input values of rule 1.20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{1}=\beta {x}_{i}\left(A\right), i=\mathrm{1,2},{R}_{1}=\beta {y}_{i-2}\left(B\right), i=\mathrm{3,4}$$\end{document}eq. no (20) denotes that it is identifying the first rule R1. Let as assume β is the Gaussian membership function, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${y}_{i}$$\end{document} is the membership values of β. Where β,21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta \left(A\right)={exp}^{-{\left(\frac{A-{\rho}_{i}}{{\sigma}_{i}}\right)}^{2}}$$\end{document} In eq. no (21), where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rho}_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma}_{i}$$\end{document} presents premise parameter set.22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{2}=\beta {x}_{i}\left(A\right)*\beta {y}_{i-2}\left(B\right)$$\end{document} For second rule R2, eq. no (22) can be used.23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{3}=\overline{{W }_{i}}=\frac{{W}_{i}}{\sum W}$$\end{document} For estimating the third rule R3, eq. no (23) can be used to analyse the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{{W }_{i}}$$\end{document}. Let us assume W is the weight.24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{4}=\overline{{W }_{i}}F=\overline{{W }_{i}}\left({l}_{i}A+{k}_{i}B+{n}_{i}\right)$$\end{document} Eq. no (24) is finding the R4 based on weight function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{{W }_{i}}F$$\end{document}. Where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{i}$$\end{document} is the consequent parameters of node i, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${n}_{i}$$\end{document} represent subset features input.25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{5}=\sum \overline{{W }_{i}}{F}_{i}$$\end{document}26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C\left(\theta \right)=R(n)\sum_{i}{\left[{y}_{T}\right]}_{x}\mathrm{log}({a}^{L}{\left({x}_{T}\right)}_{x})+(1-{\left[{y}_{T}\right]}_{x})\mathrm{lo}\mathrm{g}(1-({a}^{L}{\left({x}_{T}\right)}_{x}))$$\end{document} In eq. no (26), where yT is the label vector (Attacker/Normal), xTis the training data vector, and aL(xT) is the neural network’s output at layer L. The set of neural network parameters is represented by θ. To the attention of logical activation function the processing hidden layer are integration by constructing neural links based on MLRNN properties. The implementation of the sigmoid and hyperbolic tangent (Tanh) functions is done in the hidden layer.27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\left[\sigma (z (x,y))\right]}_{x}=\frac{1}{1+{e}^{{-z}_{k}}} , \frac{{e}^{{z}_{x}- {e}^{{-z}_{x}}}}{{e}^{{z}_{x}+{e}^{{-z}_{x}}}}$$\end{document} In eq. no (27), where z-input, k-entry count. Eq. no (28) describes a smooth maximum activation problem with gradient descent that is used to implement the output layer.28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\left[\sigma \left(z\right)\right]}_{x}=\frac{{e}^{z}x}{\sum_{a}{e}^{z}b}$$\end{document} The learning model is trained by using the ratio of the total feed forward input values to the number of input vectors which yields the average features. The average change in response value for a single unit of predictor change when the other predictor is held constant. The difference in behavior between the malicious activations and the regular transmission labels can be seen by obtaining the active output. After the IBGWO, the optimised embedding-based feature vectors are taken in as the input to the MLRNN to help in the final classification. The optimisation of the MLRNN model was done through the supervised learning strategy to maximize the classification. Empirical selection Hyperparameters (the learning rate, batch size, number of hidden layers, and number of neurons) were chosen through an initial set of experiments with the training folds of the dataset. To get fixed convergence, Adam optimizer was used, and categorical cross-entropy was taken as the loss function. Validation loss was used to stop the early termination before overfitting. The fine-tuning has been done in the scope of a 10-fold cross-validation to guarantee sound, objective performance assessment, as presented in table 2.Table 2Proposed fine tuning parameters.ParametersValuesNeurons per layer64, 128ActivationSoftmaxOptimizerAdamLearning rate0.001Batch size32Epochs50Loss functionCategorical cross entropyValidation method10-fold cross validationRegularizationEarly stopping Table 2Proposed fine tuning parameters.ParametersValuesNeurons per layer64, 128ActivationSoftmaxOptimizerAdamLearning rate0.001Batch size32Epochs50Loss functionCategorical cross entropyValidation method10-fold cross validationRegularizationEarly stopping Proposed fine tuning parameters. Results and discussion The performance coverage and dynamic topology were established to generate transmission at regular intervals under various parameter and threshold issues using the Sniffer monitor fraudulent dataset. Using a variety of conflicts settings, the Standard dataset attained was used to implement detection accuracy. A variety of simulation parameters are used to analyse the behavior of the suggested method. The representation is shown in the graphs and charts below. The results and debates of the suggested methods are presented in this volume in a descriptive manner. Additionally, dataset-based user sentiment is employed in this, with the specifications as indicated in Table 3.Table 3Simulation values for proposed method.Simulation ParametersValuesToolAnacondaEnvironmentJupyter NotebookLanguagePythonDataset NameFraudulent Website DatasetNumber of users/records3000/5000Number of classesPositive/Negative/Neutral Table 3Simulation values for proposed method.Simulation ParametersValuesToolAnacondaEnvironmentJupyter NotebookLanguagePythonDataset NameFraudulent Website DatasetNumber of users/records3000/5000Number of classesPositive/Negative/Neutral Simulation values for proposed method. Accuracy is the correct malicious activity detection based on the total number of fraudulent websites. Malicious website detection Accuracy (MWA) =\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{A}_{d}}{T\left({W}_{g}\right)}$$\end{document}, Here we assume, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C}_{d}$$\end{document} denotes correctly detect the website fraudulent and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T\left({W}_{g}\right)$$\end{document} is a total number of fraudulent website accuracy. To compare the models in a fair and unbiased manner, each of the approaches was tested with the use of one preprocessed dataset and the same training and test splits and a common 10-fold cross-validation scheme. Independent fine-tuning of model-specific hyperparameters was done on the training folds to produce the optimum performances of each method, and the evaluation metrics were calculated of the respective test folds. This comparison was done in terms of the standard performance measures such as the accuracy, precision, recall, the F-measure, error rate and the computational time. With this standardized assessment test, discrepancies in performance will be a measure of the efficacy of the models as opposed to experimental differences. Sensitivity Sensitivity is the probability that the proposed method identified fraud correctly. Sensitivity =\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{TP}^{2}V}{{TP}^{2}V+FNV}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${TP}^{2}V$$\end{document} Refers to True positive predictive values, FNV false negative predictive values Specificity The specificity is the probability that the proposed technique rightly estimates a negative. Specificity=\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{FP}^{2}V}{{FP}^{2}V+TNV}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${FP}^{2}V$$\end{document} False positive predictive values, TNV refers true negative predictive values False classification ratio The inaccuracy of the algorithm’s attack detection is indicated by the false ratio. It is measured using the below formula: False ratio (Fr) =\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{T}_{f}+{F}_{t}}{{R}_{t}}$$\end{document} Here \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${T}_{f}$$\end{document} is a classified as false, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{t}$$\end{document} is a categorize as true and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{t}$$\end{document} is a total number of user behavior. Time complexity Time complexity measures how long it takes to detect or classify. The O(n) classified at in maximum time taken for processing at in leverage point of data execution. Figure 5, defines a high impact rate level highlights outputs that detect edges of feature representation that can affect interdependencies of service level integrity rate.Fig. 5Impact of accuracy performance. Impact of accuracy performance. Fig. 5 indicates the effect of the performance of the intrusion detection classification accuracy with different services, such as 1000, 2000, and 3000. For a sample of 3000 users, the suggested MLRNN method achieved 94.3%; in comparison, the previous ReMEMBeR method achieved 78.3%, SVM achieved 81.8%, and CNN achieved 87.8%. Nevertheless, compared to alternative approaches, the suggested method yields superior results. Fig. 6 shows the MLRNN algorithm’s sensitivity performance for fraudulent detection. For three thousand users, the suggested method yields a sensitivity performance of 96%; in comparison, the results of the existing algorithms are as follows: The sensitivity performance of SVM, ReMEMBer and CNN are 76%, 81% and 85% respectively.Fig. 6Analysis of sensitivity performance. Analysis of sensitivity performance. The comparison of the suggested and earlier methods’ specificity performance study resultsare illustrated in the Fig. 7. The performance of the suggested MLRNN method is 93.8% specific. Similarly, for 3000 users, the current algorithm results for CNN, ReMEMBeR, and SVM are 78.4%, 83.6%, and 87.4% respectively, in terms of the performance of specificity.Fig. 7Analysis of specificity performance. Analysis of specificity performance. The comparison of the incorrect classification ratio performance’s effects shown in Fig. 8. Additionally, the Y-axis performance steadily declines with each approach, and the X-axis is a comparison method. In addition to CNN’s 20.7% false classification performance, the ReMEMBeR method’s 16.9%, and the SVM technique’s 12.6%, the proposed MLRNN approach obtains 1.7% false classification performance for 3000 users.Fig. 8Impact of false classification performance. Impact of false classification performance. Fig. 9 denotes temporal complexity performance of the suggested MLRNN algorithm in comparison to SVM, CNN, and ReMEMBeR. In addition to the Y-axis, which shows the time complexity performance in seconds for each approach, the X-axis in figure displays comparison methodologies. Nevertheless, compared to earlier methods, the suggested method yielded results with reduced time complexity.Fig. 9Result of time complexity performance. Result of time complexity performance. Table 4 shows the comparison analysis of proposed method with previous methods for Fraudulent Website Detection performance.Table 4Comparison analysis of proposed method with existing methods.MethodAccuracy (%)Sensitivity (%)Specificity (%)False rate (%)Key advantagesLimitationsGAN80.3181.0282.4120.45Data augmentationTraining instabilityCS-OCAN83.4683.0684.0617.8One-class learningPoor generalizationFSTL-SA IBGWO90.289.988.1710.1Few-shot learningNot web-content-orientedBiLSTM92.692.491.036.8Sequential modelingHigh computationProposed approach94.39693.81.7Optimized features and sentiment---- Table 4Comparison analysis of proposed method with existing methods.MethodAccuracy (%)Sensitivity (%)Specificity (%)False rate (%)Key advantagesLimitationsGAN80.3181.0282.4120.45Data augmentationTraining instabilityCS-OCAN83.4683.0684.0617.8One-class learningPoor generalizationFSTL-SA IBGWO90.289.988.1710.1Few-shot learningNot web-content-orientedBiLSTM92.692.491.036.8Sequential modelingHigh computationProposed approach94.39693.81.7Optimized features and sentiment---- Comparison analysis of proposed method with existing methods. Discussion Experimental results show that the proposed IBGWO-based MLRNN framework consistently outperforms traditional machine learning and deep learning approaches in detecting fraudulent websites. The performance improvement is mainly due to an effective feature optimization strategy that reduces redundancy in the high-dimensional feature space and enhances the discriminatory ability of the classifier. By integrating A-RFE and IBGWO optimization, the model selects the optimal subset of emotion recognition and structural features, achieving superior classification performance. Accuracy analysis results show that the proposed method achieves 94.3% accuracy, outperforming CNN, SVM, and ReMEMBeR models. This improvement indicates that the optimized feature subset allows MLRNN to capture better more complex patterns associated with fraudulent behavior. Similarly, the 96% sensitivity highlights the model’s impressive ability to correctly identify phishing websites, which is crucial for minimizing security risks in real-world applications. High sensitivity means fewer fraudulent websites escape detection, providing better protection for users. Specificity analysis further confirmed the robustness of the proposed approach, achieving 93.8% specificity, significantly higher than that of existing methods. The results show that the proposed framework effectively reduces false alarms by correctly classifying legitimate websites. Furthermore, the low misclassification rate of 1.7% demonstrates the proposed system’s ability to distinguish between legitimate and phishing websites, thereby improving its reliability and usability. The time-complexity evaluation shows that the proposed MLRNN framework requires less computational time than the CNN and ReMEMBeR methods. This time complexity reduction is mainly due to the optimized feature subset, which reduces the dimensionality of the input data and accelerates the training and inference processes. Therefore, the proposed method is suitable for practical deployment where timely detection is essential. The comparative analysis shown in Table 4 further validates the effectiveness of the proposed approach against state-of-the-art techniques such as GAN, CS-OCAN, BiLSTM, and FSTL-SA based models. Although existing approaches have shown reasonable performance, they suffer from limitations such as training instability, low generalization, high computational cost, and lack of web content orientation. In contrast, the proposed IBGWO-MLRNN framework achieves the best overall performance by balancing accuracy, sensitivity, specificity, and computational efficiency. Despite its good performance, the proposed approach has certain limitations. First, some functionality is derived from external services, which can introduce dependency and latency issues into a real-time phishing detection system. Second, while the evaluation is performed on a balanced benchmark dataset, real-world phishing scenarios are often highly imbalanced, which may affect generalization performance. Furthermore, although the MLRNN model has reduced time complexity compared to other deep learning methods, further optimization may be required for deployment in large-scale real-time environments. Conclusion To conclude, Multi-Layer Recognition Neural Networks (MLRNNs) have shown to be a sophisticated and successful technique in the realm of network security for identifying phony websites. Because of their remarkable capacity to recognize intricate patterns and abnormalities in datasets, MLRNNs—which are composed of linked layers of neurons—are ideally adapted to the intricate and dynamic nature of online fraud activity. In order to identify fake websites, MLNN design meticulously incorporates variables like Uniform Resource Locator structure, Domain Age, IP address reputation, Secure Socket Layer certificate data, and the content characteristics. Input layers for pertinent features, hidden layers for nonlinear transformations, and output layers for binary classification (fraudulent or legal) are the typical components of neural network topologies. MLRNN gains the ability to recognize patterns that can point to fraudulent activity through an iterative training procedure that makes use of optimization strategies. MLRNN and IBGWO to combine the best features of reduced complexity and implementation requirements make them difficult to scale. Machine learning models, especially neural networks, hold the promise of detecting fraud by extracting complex patterns from large volumes of data. Effective training requires large volumes of labeled data and processing. The proposed approach effectively addresses key challenges in fraudulent website detection by leveraging sentiment-aware analysis and Improved Binary Grey Wolf Optimization for robust feature selection.Although the proposed IBGWO–MLRNN framework demonstrates improved performance in fraudulent website detection, several enhancements can be explored in future work. The model can be extended to support real-time detection by incorporating online learning mechanisms. Additionally, integrating transformer-based language models may further improve sentiment representation from complex web content. Based on the experimental findings, several avenues for future research can be identified. First, the strong performance gains achieved through optimized feature selection suggest exploring adaptive and dynamic feature optimization strategies for evolving fraudulent patterns. Second, the effectiveness of sentiment-based features indicates potential benefits from integrating advanced language models, such as transformer-based architectures, to capture deeper contextual semantics. Third, future work may focus on extending the proposed framework to real-time and large-scale environments using online learning mechanisms. Finally, evaluating the model on multilingual and cross-domain datasets could further enhance its robustness and generalization capability. Future studies may also investigate adaptive optimisation strategies to handle evolving phishing patterns and evaluate the framework on larger, multilingual, and cross-domain datasets to enhance generalization.