Wednesday, November 18, 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

Dual-Layer Machine Learning Framework Detects Unknown Cyberattacks While Reducing False Alarms

November 18, 2026
in Science News
Blake Davidson
By Blake Davidson Scienmag Editorial Profile - Data Science
Reading Time: 6 mins read
0
Dual-Layer Machine Learning Framework Detects Unknown Cyberattacks While Reducing False Alarms

Dual-Layer Machine Learning Framework Detects Unknown Cyberattacks While Reducing False Alarms

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

A dual-layer artificial intelligence framework combining supervised decision ensembles with unsupervised deep neural networks has demonstrated high accuracy in identifying unexpected cyber threats while reducing false alarms in network traffic. Published in the Journal on Information Security, the empirical study evaluated a hybrid intrusion detection architecture coupling a Random Forest classifier with a deep Autoencoder to inspect digital packets in real time. Across evaluations on two benchmark datasets, the proposed two-stage methodology achieved a peak F1-score of 95.90 percent on the CICIDS 2018 dataset and 99.72 percent on the SCVIC-APT dataset. Crucially, the layered strategy addressed a major operational challenge in cybersecurity: the vulnerability of conventional supervised classifiers to unmodeled attack signatures alongside the burden of false alarms. By extracting data signatures across complementary dimensions, the hybrid system recorded a false positive rate of only 1.81 percent on the CICIDS 2018 benchmark, outperforming individual baseline algorithms and standard single-stage detection systems.

Modern network security infrastructure relies on automated intrusion detection systems to inspect packet streams, detect unauthorized activity, and defend systems against malicious network intrusions. However, security operations centers face technical bottlenecks when deploying traditional machine learning models into operational enterprise environments. Conventional supervised classifiers depend on historical datasets containing labeled records of known benign traffic and specific attack categories, including denial-of-service, brute-force attempts, and botnet propagation. When an adversary deploys an unmodeled exploit, introduces a novel assault method, or modifies an established attack vector, standard supervised classifiers frequently fail to generalize beyond their explicit training distributions. Under standard deployment parameters, these models misclassify unprecedented malicious actions as ordinary background communication, generating low detection rates for unmodeled threats. Conversely, models configured with overly aggressive anomaly thresholds trigger excessive false alarms on non-malicious variations, overwhelming human operators and increasing the risk that authentic cyber intrusions will be overlooked during active monitoring.

To overcome these limitations, the research team designed an integrated two-stage detection pipeline altering how ensemble models interpret classification uncertainty during initial traffic triage. In the primary operational phase, incoming network records are processed by an ensemble of Decision Trees constructed through bootstrap aggregation and feature randomization, forming a Random Forest. Rather than utilizing standard deterministic prediction functions assigning samples to the category bearing the highest overall vote, the architecture computes continuous probability estimates through the scikit-learn predict_proba function. Individual decision trees establish class probabilities by tallying training samples within leaf nodes, which the global ensemble averages into unified probability metrics. The researchers established an empirical decision boundary, designated as threshold T1, calculated from the 90th percentile of probability scores observed across normal validation traffic. Any incoming transmission generating an attack probability exceeding this calibrated threshold is categorized as an assault, enabling the system to intercept atypical threats exhibiting intermediate probability values.

While lowering the initial classification threshold enables the Random Forest to capture unidentified attack types, it inevitably flags ambiguous benign packets whose feature distributions overlap with hostile behavior. To resolve this precision deficit without discarding true threat alerts, researchers introduced a secondary validation layer driven by an unsupervised deep Autoencoder. Deep autoencoders function through symmetrical neural network architectures comprising encoder and decoder components separated by compressed latent representations. In the evaluated implementation, the encoder compresses input features through hidden layers into a lower-dimensional latent space, after which the decoder attempts to reconstruct the original input vector. Crucially, the autoencoder is trained exclusively on validated normal communication flows, learning to reconstruct ordinary network behaviors with negligible distortion. Because attack signatures never participate in training autoencoder weights, anomalous flows generate substantial reconstruction discrepancies. The degree of reconstruction failure is formally quantified through Mean Square Error loss calculations, providing a sensitive unsupervised indicator of packet irregularity.

The core operational innovation of the architecture lies in how these two algorithmic paradigms collaborate within a structured decision workflow. When an incoming network record is evaluated, it first undergoes screening by the probabilistic Random Forest; if its predicted attack probability does not exceed threshold T1, the system designates it as benign. Conversely, network samples flagged by the Random Forest as potential threats are redirected to the second-stage Autoencoder for secondary verification rather than triggering an alarm. The Autoencoder evaluates the sample and computes its Mean Square Error reconstruction value, comparing it against hyperparameter T2. Configured at the 75th percentile of reconstruction errors observed on normal validation data, T2 serves as an anomalous error ceiling. If a suspected attack sample exhibits a reconstruction error smaller than T2, the system recognizes that the flow conforms to normal traffic geometry and reclassifies it as benign, successfully extinguishing what would otherwise have manifested as an erroneous false positive alert.

To validate their framework under conditions mimicking zero-day deployments, the authors conducted controlled experiments using two established intrusion benchmarks: CICIDS 2018 and SCVIC-APT. The CICIDS 2018 dataset contains six attack classes alongside benign activity: brute force, denial-of-service, distributed denial-of-service, botnet, infiltration, and web attacks. The SCVIC-APT benchmark documents advanced persistent threats across four categories. Both datasets were constructed using 43 NetFlow v9 attributes per record. The pre-processing workflow removed source IP addresses, handled numerical values with logarithmic transformations, encoded categorical factors via one-hot encoding, and normalized continuous values into a zero-to-one range. Across sampled distributions encompassing approximately 300 records for CICIDS 2018 and 200 records for SCVIC-APT, data was split in a four-to-one training-to-testing ratio, with 10 percent of training data reserved for validation. Crucially, researchers simulated unknown attacks by excluding one attack category from training, running five trials per category across 30 CICIDS 2018 experiments and 20 SCVIC-APT experiments.

Empirical comparisons against established baselines demonstrated marked advantages for the hybrid methodology. The researchers evaluated the framework against Decision Trees, Logistic Regression, standard Random Forest, Deep Neural Networks, Isolation Forests, One-Class Support Vector Machines, and a standalone Autoencoder. When denial-of-service attacks were held out as unknown threats in the CICIDS 2018 dataset, the conventional Random Forest classifier mistakenly labeled 64.8 percent of attack samples as normal traffic. In contrast, the probability-guided Random Forest misclassified merely 1.8 percent of attack samples as normal, proving that probability thresholding substantially improves sensitivity to unseen threats. However, this sensitivity heightened false alarms when used in isolation. By appending the Autoencoder verification module, the dual-layer system achieved an exceptional balance of precision and recall, reducing false positive rates to 1.81 percent while delivering overall F1-scores of 95.90 percent on CICIDS 2018 and 99.72 percent on SCVIC-APT, surpassing conventional supervised algorithms and isolation techniques.

Performance evaluations of individual machine learning subcomponents confirmed the mathematical stability of the pipeline. Tested as an independent anomaly detector, the deep Autoencoder achieved 97.77 percent accuracy, 99.67 percent precision, 100 percent recall, and an F1-score of 99.50 percent, demonstrating high capability in identifying abnormal patterns. Receiver Operating Characteristic analysis yielded an Area Under the Curve of 0.9708 for the model, illustrating separation between normal traffic and malicious flows across diverse decision thresholds. Specifically, the curve documented a 74 percent true positive rate at a stringent false positive rate of merely 2 percent. Reconstruction error histograms revealed a pronounced gap between normal distributions concentrated at the lower boundary and anomalous samples scattered across elevated error ranges. When assessing overall classification outcomes across traffic classes, the complete intrusion detection framework achieved a general test accuracy of 94.62 percent, exhibiting near-perfect detection for Botnet, DDoS, and benign streams while highlighting misclassification in Infiltration and Web attack categories.

The investigation subjected the architecture to overfitting examinations and parameter sensitivity analyses to ensure operational reliability across network environments. Tenfold stratified cross-validation executed on the Random Forest classifier demonstrated consistency, recording a standard deviation of only 0.10 percent across folds alongside an F1-score range between 98.1 percent and 98.4 percent. Furthermore, empirical learning curves tracking performance as training volume expanded from 10 percent to 100 percent revealed convergence between training and validation trajectories. The final gap between training performance at 98.7 percent and validation performance at 98.3 percent measured just 0.4 percent, with both curves plateauing after reaching 70 percent of training data. Systematic sensitivity testing revealed that model performance peaked at 98.2 percent F1 when threshold T1 occupied the 90th percentile. Similarly, the T2 reconstruction threshold achieved optimal performance of 98.4 percent F1 at the chosen 75th percentile, with statistical testing confirming that extreme threshold variations caused significant performance drops exceeding 0.5 percent.

Beyond theoretical accuracy, the authors analyzed computational scalability and real-time execution capabilities required for live production deployments. The system incorporates an operational pipeline featuring continuous packet capture, automated data filtering, and log generation integrated via a Flask bridge with a graphical dashboard. In benchmark evaluations, the hybrid model completed training on the CICIDS 2018 dataset in approximately eight minutes, representing a manageable overhead for periodic retraining. During inference, the hybrid system processed 5,208 samples per second, with individual latency measuring 0.192 milliseconds under unbatched execution. Under batch processing, inference latency dropped to 0.098 milliseconds at batch size 64 and reached 0.076 milliseconds at batch size 256, achieving a 60 percent throughput improvement. Coupled with an inference memory footprint of 223 megabytes, the framework demonstrates practicality for enterprise networks requiring sustained processing rates between 500 and 1,000 samples per second. However, the current framework operates as a binary classifier achieving 98.2 percent accuracy, requiring future development of one-versus-rest multi-class models to provide granular threat identification for incident response teams.

Subject of Research: Dual-layer network intrusion detection combining random forest and autoencoder machine learning models

Article Title: A dual-layer AI-based approach for real-time network intrusion detection using random forest and autoencoder

Article References: Khemani, B., Sapkale, S., Buchade, T., Singh, S., Asthana, S., Patil, S., & Wankhede, K. (2026). A dual-layer AI-based approach for real-time network intrusion detection using random forest and autoencoder. Journal on Information Security, 2026(1), Article 23. https://doi.org/10.1186/s13635-026-00239-5

Image Credits: AI Generated

DOI: 10.1186/s13635-026-00239-5

Keywords: network intrusion detection, machine learning, random forest, autoencoder, deep learning, anomaly detection, cybersecurity, CICIDS 2018, SCVIC-APT, false positive rate, reconstruction error, decision trees, network traffic analysis, predictive probability, intrusion detection system, mean square error, binary classification

Cite Scienmag News

Blake Davidson. (November 18, 2026). Dual-Layer Machine Learning Framework Detects Unknown Cyberattacks While Reducing False Alarms. Scienmag. https://scienmag.com/dual-layer-machine-learning-framework-detects-unknown-cyberattacks-while-reducing-false-alarms/

Blake Davidson. "Dual-Layer Machine Learning Framework Detects Unknown Cyberattacks While Reducing False Alarms." Scienmag, 18 November 2026, https://scienmag.com/dual-layer-machine-learning-framework-detects-unknown-cyberattacks-while-reducing-false-alarms/. Accessed 18 November 2026.

Blake Davidson. "Dual-Layer Machine Learning Framework Detects Unknown Cyberattacks While Reducing False Alarms." Scienmag. November 18, 2026. https://scienmag.com/dual-layer-machine-learning-framework-detects-unknown-cyberattacks-while-reducing-false-alarms/

Tags: anomaly detectionautoencoderbinary classificationCICIDS 2018cybersecuritydecision treesdeep learningfalse positive rateintrusion detection systemMachine learningmean square errornetwork intrusion detectionnetwork traffic analysispredictive probabilityRandom Forestreconstruction errorSCVIC-APT
Share26Tweet16
Previous Post

Deep Learning Model YOLOv8n-seg Improves Laser Experiment Segmentation Efficiency and Accuracy

Next Post

Spring Precipitation Is Negatively Related to Earlywood Density in Picea Glehnii

Related Posts

Decitabine and Selinexor Before Stem Cell Transplant Show Promise in High-Risk MDS
Cancer

Decitabine and Selinexor Before Stem Cell Transplant Show Promise in High-Risk MDS

November 18, 2026
Cardiovascular Disease and Cancer Drive Life Expectancy Gaps in Fiji
Medicine

Cardiovascular Disease and Cancer Drive Life Expectancy Gaps in Fiji

November 18, 2026
Advanced Catalytic Systems Accelerate Sustainable Breakdown of Hazardous Industrial Wastewater Pollutants
Chemistry

Advanced Catalytic Systems Accelerate Sustainable Breakdown of Hazardous Industrial Wastewater Pollutants

November 18, 2026
Blockchain-Enabled Federated Learning System Optimizes Distributed Resource Allocation
Technology and Engineering

Blockchain-Enabled Federated Learning System Optimizes Distributed Resource Allocation

November 18, 2026
Czech Dietitians Rely on Predictive Equations Despite Critical Accuracy Knowledge Gaps
Medicine

Czech Dietitians Rely on Predictive Equations Despite Critical Accuracy Knowledge Gaps

November 18, 2026
Applying Ethephon to Peanut Plants Reduces Seed Dormancy and Improves Vigor Testing
Science News

Applying Ethephon to Peanut Plants Reduces Seed Dormancy and Improves Vigor Testing

November 18, 2026
Next Post
Spring Precipitation Is Negatively Related to Earlywood Density in Picea Glehnii

Spring Precipitation Is Negatively Related to Earlywood Density in Picea Glehnii

  • 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

  • Decitabine and Selinexor Before Stem Cell Transplant Show Promise in High-Risk MDS
  • Cardiovascular Disease and Cancer Drive Life Expectancy Gaps in Fiji
  • Advanced Catalytic Systems Accelerate Sustainable Breakdown of Hazardous Industrial Wastewater Pollutants
  • Blockchain-Enabled Federated Learning System Optimizes Distributed Resource Allocation

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