Thursday, November 19, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News

Ensemble Machine Learning Model Detects Phishing Websites With High Accuracy and Efficiency

November 19, 2026
in Science News
Blake Davidson
By Blake Davidson Scienmag Editorial Profile - Data Science
Reading Time: 6 mins read
0
Ensemble Machine Learning Model Detects Phishing Websites With High Accuracy and Efficiency

Ensemble Machine Learning Model Detects Phishing Websites With High Accuracy and Efficiency

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

Computer scientists have designed and validated EnLeM, an ensemble learning architecture that combines three supervised machine learning algorithms to detect fraudulent phishing websites with high classification accuracy while significantly reducing computational runtime. Phishing attacks function as social engineering campaigns where deceptive actors masquerade as legitimate entities to illicitly acquire sensitive credentials, banking information, user identifiers, and passwords through deceptive emails, telephone calls, text messages, and fabricated web platforms. Traditional defense mechanisms have historically relied on heuristics, blacklisting databases containing known malicious signatures, or visual similarity analyses that compare page formatting and logos against verified repositories. However, heuristic techniques often involve complex and costly implementations, signature-based blacklists fail against newly engineered URLs, and visual comparison tools impose heavy processing overhead. By coordinating three complementary classifiers under a hard-voting mechanism and integrating univariate feature selection, the newly demonstrated framework provides an effective, lightweight alternative capable of operating efficiently without the severe resource demands associated with deep neural network pipelines.

The research team assembled the EnLeM architecture by integrating three well-established supervised learning classifiers: a decision tree, a random forest, and a k-nearest neighbor algorithm. Each constituent classifier contributes distinct operational advantages to the overall detection system. Decision trees generate human-interpretable conditional rules from continuous and categorical variables, although individual trees can exhibit variance and instability. Random forests mitigate these tendencies by generating ensembles of decision trees trained across bootstrapped samples of the dataset, effectively reducing variance and resisting overfitting while identifying key attribute importance. Meanwhile, the k-nearest neighbor algorithm provides localized instance-based evaluations by identifying the nearest data points across feature spaces. To synthesize these base learners, the system applies a hard-voting protocol where each individual algorithm outputs a binary prediction classifying an unlabeled webpage as either legitimate or phishing. The overall ensemble adopts the categorical prediction backed by the majority of the constituent classifiers. To ensure deterministic operations during potential three-way ties, the architecture incorporates a predefined rule that defaults to the decision of the random forest, which exhibited the highest standalone predictive accuracy during preliminary benchmark testing.

To evaluate the system under rigorous benchmark standards, the researchers conducted experiments using the public UCI phishing website repository, which comprises 11,055 total webpage instances consisting of 6,157 confirmed legitimate sites and 4,898 verified phishing URLs. Realizing consistent performance across supervised classification models requires comprehensive data preprocessing to eliminate structural biases. The researchers implemented standard data cleaning protocols to remove statistical outliers, eliminate duplicate records, handle missing values, and normalize feature measurements into uniform ranges. Normalization prevents attributes with large numerical scales from dominating model optimization while ensuring that each attribute maintains an equal numerical contribution before evaluation. Categorical targets were converted into binary numeric labels, where a value of zero designated a legitimate webpage and a value of one represented a fraudulent phishing site. While initial feature extraction yielded 30 distinct attributes characterizing URL structures and page properties, extraneous and redundant variables often trigger the curse of dimensionality, creating unnecessary latency and elevating false-positive rates in automated security applications.

To overcome dimensionality bottlenecks without compromising classification fidelity, the researchers incorporated an automated univariate feature selection protocol based on Mutual Information. Mutual Information evaluates the shared statistical dependency between individual independent variables and the target classification label, measuring the difference between the marginal entropy of an attribute and its conditional entropy relative to the target class. Using the SelectKBest method, the pipeline ranked all 30 available attributes and isolated the top 20 most informative features for subsequent model evaluation. Crucially, the researchers embedded this feature selection procedure directly inside each fold of a 10-fold cross-validation scheme to prevent data leakage. In each split, the mutual information scores and feature subsets were fitted solely on the designated training fold and subsequently mapped onto the validation fold. This methodological safeguard ensured that validation observations remained entirely unseen during the dimensionality reduction phase, preventing artificial performance inflation and maintaining rigorous evaluation integrity.

The researchers also conducted a targeted ablation study to verify how varying the feature selection threshold influenced model stability and operational speed. By systematically evaluating performance across subsets containing 10, 15, 20, 25, and 30 features, the investigators observed that EnLeM exhibited remarkable resilience to dimensionality reductions. Classification accuracy and F1-score metrics remained highly stable, varying by less than 0.7% across all tested feature thresholds. At the same time, reducing feature counts delivered clear computational benefits, decreasing overall execution time from 10.01 seconds down to 8.72 seconds as redundant attributes were pruned. This experimental finding demonstrated that the Mutual Information framework successfully isolates salient phishing indicators while discarding noisy or irrelevant features that otherwise consume memory bandwidth and processing cycles during real-time traffic analysis.

To ensure reproducible and standardized evaluations, the hyperparameters governing the base learners were systematically tuned using grid search procedures coupled with 10-fold cross-validation rather than arbitrary manual assignments. The parameter exploration assessed decision tree and random forest maximum tree depths across values of 3, 5, 7, and 10, while testing random forest ensemble sizes across 50, 100, and 200 estimators. For the k-nearest neighbor classifier, the grid search examined neighborhood values of 3, 5, 7, and 9 utilizing both Euclidean and Manhattan distance metrics. The final operational settings were selected strictly based on the configurations that maximized average validation accuracy across the cross-validation splits. The experimental computing environment utilized an NVIDIA Tesla K80 graphical processing unit paired with an Intel Xeon hyperthreaded processor running at 2.3 GHz, establishing a controlled hardware baseline for comparative timing and throughput assessments.

Across initial benchmark evaluations utilizing the complete 30-feature dataset without feature selection, EnLeM demonstrated superior detection capabilities compared to standard machine learning and deep learning architectures. Evaluated under 10-fold cross-validation, EnLeM attained an overall accuracy of 97.21%, a precision of 97.06%, a recall of 97.97%, and an F1-score of 97.51%, reflecting an empirical error rate of just 0.0279. In direct comparison, individual machine learning baselines exhibited lower overall classification performance: standalone decision trees achieved 96.75% accuracy, random forests recorded 92.37%, AdaBoost reached 93.21%, k-nearest neighbor logged 91.71%, and support vector machines scored 94.74%. When contrasted with complex deep learning baselines, EnLeM also outperformed a one-dimensional convolutional neural network, which achieved 96.75% accuracy and an F1-score of 96.24%, as well as a long short-term memory recurrent neural network, which posted 90.84% accuracy and a 91.75% F1-score. Because the evaluation dataset contains an asymmetric distribution of legitimate and phishing pages, the harmonic mean captured by the F1-score provided a robust measure confirming that EnLeM balanced false-positive and false-negative errors effectively.

When evaluated in conjunction with the 20-feature Mutual Information selection pipeline, EnLeM maintained its performance leadership while drastically widening its efficiency advantage over deep neural models. Operating on the refined top-20 feature subset, EnLeM achieved an accuracy of 96.26%, a precision of 96.36%, a recall of 97.96%, and an F1-score of 96.66%, outperforming all traditional machine learning baselines, which clustered between 91.26% and 94.35% accuracy. In stark contrast, the performance of the one-dimensional convolutional neural network collapsed under the reduced feature regime, dropping sharply to an accuracy of 78.06% and recording the lowest F1-score across all tested architectures. This deterioration aligns with broader empirical evidence indicating that deep convolutional networks struggle to extract effective hierarchical representations when tabular inputs are compressed through univariate selection techniques, highlighting the inherent fragility of deep architectures in restricted feature environments.

A formal statistical assessment confirmed that the operational advantages of dimensionality reduction were achieved without any statistically significant degradation in detection quality. A paired t-test comparing EnLeM’s performance across accuracy, precision, recall, and F1-score before and after feature selection yielded a p-value of 0.1353. Because this value sits comfortably above the conventional 0.05 significance threshold, the minor numerical differences are considered statistically indistinguishable. However, in terms of execution latency, feature reduction trimmed processing time by 1.29 seconds, representing an approximate 13% improvement in execution speed. While simple standalone learners such as basic decision trees exhibited lower raw runtimes, their standalone predictive accuracy remained inferior, and resource-heavy models like long short-term memory networks suffered extensive training delays due to memory bandwidth constraints. The combination of majority hard voting and mutual information pruning allowed EnLeM to secure optimal predictive reliability while accelerating execution throughput.

The authors contextualized their findings within the current scope of cybersecurity research while outlining clear boundaries and prospective technical expansions. The present study evaluated EnLeM on a centralized repository of 11,055 pre-extracted URL features, meaning that real-world deployment across distributed gateway infrastructure, live browser extensions, and evolving evasion techniques will require continuous evaluation. Deep neural networks remain computationally prohibitive for lightweight endpoint security, whereas EnLeM proves that combining interpretable base learners through principled ensemble rules yields resilient, high-speed protection against malicious web platforms. The research team noted that future investigations will focus on adapting EnLeM to decentralized and privacy-preserving architectures such as federated learning, exploring non-linear dimensionality reduction methods like autoencoders, and benchmarking the ensemble against larger, dynamically captured global threat streams.

Subject of Research: Machine learning ensemble model for phishing website detection

Article Title: EnLeM: ensemble learning-based model to detect phishing websites

Article References: Yeasmin, M. N., Refat, M. A. R., Singh, B. C., Alom, Z., Aung, Z., & Azim, M. (2025). EnLeM: ensemble learning-based model to detect phishing websites. EURASIP Journal on Information Security, 2026(1), Article 1. https://doi.org/10.1186/s13635-025-00219-1

Image Credits: AI Generated

DOI: 10.1186/s13635-025-00219-1

Keywords: phishing detection, ensemble learning, cybersecurity, machine learning, decision tree, random forest, k-nearest neighbor, hard voting, feature selection, mutual information, web security, UCI dataset, deep learning benchmark, classification algorithms, data preprocessing, social engineering

Cite Scienmag News

Blake Davidson. (November 19, 2026). Ensemble Machine Learning Model Detects Phishing Websites With High Accuracy and Efficiency. Scienmag. https://scienmag.com/ensemble-machine-learning-model-detects-phishing-websites-with-high-accuracy-and-efficiency/

Blake Davidson. "Ensemble Machine Learning Model Detects Phishing Websites With High Accuracy and Efficiency." Scienmag, 19 November 2026, https://scienmag.com/ensemble-machine-learning-model-detects-phishing-websites-with-high-accuracy-and-efficiency/. Accessed 19 November 2026.

Blake Davidson. "Ensemble Machine Learning Model Detects Phishing Websites With High Accuracy and Efficiency." Scienmag. November 19, 2026. https://scienmag.com/ensemble-machine-learning-model-detects-phishing-websites-with-high-accuracy-and-efficiency/

Tags: classification algorithmscybersecuritydata preprocessingdecision treedeep learning benchmarkensemble learningfeature selectionhard votingK-Nearest NeighborMachine learningmutual informationphishing detectionRandom Forestsocial engineeringUCI datasetweb security
Share26Tweet16
Previous Post

Review Examines Emerging Biomarker-Guided Therapies for Platinum-Resistant Ovarian Cancer

Next Post

Curcumin Analogue FLLL32 Triggers Ferroptotic Cell Death in Colorectal Cancer Models

Related Posts

Spatial Cell Interactions Shape Immune Escape in Hematological Malignancies Review Finds
Cancer

Spatial Cell Interactions Shape Immune Escape in Hematological Malignancies Review Finds

November 19, 2026
Kidney Xenotransplantation Proposed as Investigational Rescue Strategy Benchmarked Against High-KDPI Allografts
Medicine

Kidney Xenotransplantation Proposed as Investigational Rescue Strategy Benchmarked Against High-KDPI Allografts

November 19, 2026
Extracellular Vesicles Deliver Antimicrobial Peptides to Combat Rice Blast Disease
Science News

Extracellular Vesicles Deliver Antimicrobial Peptides to Combat Rice Blast Disease

November 19, 2026
Family Discussions and Spiritual Well-Being Shape Organ Donation Attitudes in Iran
Medicine

Family Discussions and Spiritual Well-Being Shape Organ Donation Attitudes in Iran

November 19, 2026
Most Saqqez Hospital Nurses Demonstrated Moderate Adherence to Professional Ethical Principles
Medicine

Most Saqqez Hospital Nurses Demonstrated Moderate Adherence to Professional Ethical Principles

November 19, 2026
Interviews With Cheating Contractors Reveal Hidden Homework Markets and Countermeasures
Science News

Interviews With Cheating Contractors Reveal Hidden Homework Markets and Countermeasures

November 19, 2026
Next Post
Curcumin Analogue FLLL32 Triggers Ferroptotic Cell Death in Colorectal Cancer Models

Curcumin Analogue FLLL32 Triggers Ferroptotic Cell Death in Colorectal Cancer Models

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Spatial Cell Interactions Shape Immune Escape in Hematological Malignancies Review Finds
  • Kidney Xenotransplantation Proposed as Investigational Rescue Strategy Benchmarked Against High-KDPI Allografts
  • Extracellular Vesicles Deliver Antimicrobial Peptides to Combat Rice Blast Disease
  • Family Discussions and Spiritual Well-Being Shape Organ Donation Attitudes in Iran

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading