• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Wednesday, November 18, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

Generative Adversarial Networks and Deep Learning Bolster Internet of Things Security

by
November 13, 2026
in Technology
Reading Time: 6 mins read
0
Generative Adversarial Networks and Deep Learning Bolster Internet of Things Security

Generative Adversarial Networks and Deep Learning Bolster Internet of Things Security

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A research investigation published in Discover Artificial Intelligence has demonstrated that pairing long short-term memory recurrent neural networks with generative adversarial network data augmentation significantly improves the detection of rare cyberattacks within Internet of Things environments. Modern interconnected ecosystems face mounting security vulnerabilities as rapid infrastructure growth creates immense volumes of network traffic that outpace traditional defensive tools. When security systems rely on standard anomaly detection algorithms, severely skewed class distributions often cause machine learning classifiers to prioritize dominant traffic patterns while ignoring infrequent but severe exploits. To address this persistent limitation, researchers Bhawana Sharma, Lokesh Sharma, and Chhagan Lal engineered an intrusion detection architecture that extracts distinctive temporal features, filters redundant data, and synthesizes realistic instances of underrepresented network attacks. Evaluated against the standard benchmark UNSW-NB15 dataset, the resulting generative-augmented deep learning model reached an overall classification accuracy of 92 percent, markedly exceeding the performance of baseline architectures while sharply elevating detection precision for elusive malicious behaviors.

The expansion of Internet of Things deployments across diverse operational fields has introduced critical vulnerabilities that stem directly from the immense scale of interconnected devices. The relentless escalation in network traffic yields vast quantities of data, introducing substantial technical obstacles regarding high-volume data storage and computational management. Network intrusion detection systems are broadly separated into two operational strategies based on their underlying detection methodologies: signature-based mechanisms and anomaly-based frameworks. Signature-based systems compare inbound activity against repositories of previously documented attack patterns, which renders them incapable of identifying newly engineered threats while burdening networks with finite storage capacity constraints. In contrast, anomaly-based systems identify unauthorized disruptions by discovering statistical outliers and deviations from ordinary operational baselines, thereby maintaining the capability to recognize novel or unknown exploits. However, conventional anomaly detection remains susceptible to elevated rates of false positive classifications, creating an urgent technical demand for robust computational models that can process vast datasets and manage intensive computational workflows without sacrificing diagnostic precision.

To overcome the architectural shortcomings of traditional cybersecurity methods, modern research increasingly relies on deep learning architectures capable of processing extensive datasets and executing complex computational transformations. Sequential data transmission in networked environments particularly benefits from recurrent neural networks, which maintain internal memory loops that pass previous hidden layer states into subsequent computational steps. Despite their utility for sequential processing, conventional recurrent neural networks struggle with the vanishing gradient problem, where shrinking numerical updates during backpropagation prevent weights from converging effectively during extended sequences. Long short-term memory networks solve this failure mode by integrating specialized gating cells that regulate the storage, passage, and elimination of information across time. Inside an LSTM cell, forget gates determine what past information should be discarded using sigmoid activation functions, input gates determine which newly arriving data values warrant storage through combined sigmoid and hyperbolic tangent operations, and output gates regulate the final hidden state passed to adjacent layers, ensuring dependable retention across complex traffic flows.

A primary bottleneck in training deep intrusion detection systems lies in severe class imbalance within empirical network traffic records. In benchmark datasets, standard benign communication vastly outnumbers malicious records, and specific attack variations appear far less frequently than common exploit scripts. Under such uneven distributions, learning models develop systematic biases toward majority classes, generating inflated aggregate accuracy scores while failing to identify critical minority intrusions. While basic data resampling offers a theoretical remedy, conventional strategies introduce substantial structural compromises. Simple oversampling duplicates existing minority observations, creating an acute hazard of model overfitting, whereas random undersampling discards authentic records from majority distributions, causing the permanent loss of vital operational information. Synthetic minority oversampling techniques partially alleviate these concerns by executing k-nearest neighbor interpolation, but generating points strictly along linear segments between neighboring coordinates can create artificial noise and exacerbate overlapping decision boundaries, ultimately degrading classifier reliability under complex operational conditions.

To overcome the limitations of classical synthetic sampling, the research team implemented a generative adversarial network architecture specifically tailored to synthesize realistic minority attack instances. Unlike linear interpolation, generative frameworks engage two distinct deep neural networks in an adversarial min-max optimization game governed by opposing mathematical objectives. The generator network ingests random noise distributions and attempts to generate synthetic data samples whose statistical properties match the underlying distribution of genuine network traffic, aiming to minimize the probability that its generated instances are identified as artificial. Concurrently, the discriminator network evaluates mixed inputs of genuine and synthetic records, aiming to maximize its diagnostic accuracy by assigning authentic records a value of one and generated records a value of zero. As training progresses across iterations using binary cross-entropy loss and the Adam optimization algorithm, the discriminator’s feedback continually drives the generator toward producing high-fidelity synthetic feature profiles, establishing an optimal balance where synthetic records accurately enrich minority classes without distorting genuine boundary conditions.

Empirical evaluation of the framework was conducted using the comprehensive UNSW-NB15 benchmark dataset, created at the Australian Centre for Cyber Security. The repository comprises 257,673 network records containing 43 distinct features, of which 39 possess numerical properties and four represent categorical attributes, alongside a final supervisory ground-truth label. Within the overall corpus, 93,000 records capture normal network behavior, whereas 164,673 entries encompass nine distinct cyberattack categories, including denial of service, exploits, fuzzers, generic attacks, analysis, worms, shellcode, backdoors, and reconnaissance. For experimental execution, the investigators isolated five primary operational categories consisting of normal network flows alongside four dominant attack categories: generic, exploits, denial of service, and fuzzers. Due to the high device density characteristic of modern deployments, network infrastructures remain exceptionally vulnerable to denial of service incursions, which overwhelm targeted endpoints with repetitive requests to induce system failure, as well as fuzzer attacks, which feed invalid data sequences into hardware to trigger buffer overflows and operational collapse.

To prepare raw network traffic for neural processing, the researchers executed a systematic multi-stage preprocessing protocol. Categorical attributes including attack category, service, state, and protocol—encompassing 10, 13, 11, and 133 unique identifiers respectively—were systematically mapped into discrete numerical indicators via label encoding. To prevent variables with broad numerical distributions from disproportionately dominating optimization trajectories, Min-Max normalization was subsequently applied, linearly rescaling every feature value into a standardized interval bounded between zero and one. Following normalization, feature dimensionality was curtailed using a Pearson correlation filter method to eliminate redundant indicators that elevate computational complexity without delivering diagnostic value. By calculating bivariate correlation coefficients across all attributes against an established threshold cutoff of 0.95, the team identified highly collinear variable pairs. Whenever two features exhibited mutual correlation values exceeding this 0.95 threshold, one variable was excluded from subsequent stages, effectively streamlining the feature space and reducing unnecessary model complexity prior to deep neural training.

The experimental configuration assessed three distinct methodological paradigms to quantify the operational impact of data augmentation on intrusion detection capabilities. The initial baseline implementation trained an unaugmented long short-term memory network directly upon the selected UNSW-NB15 records using a 75 percent training and 25 percent testing distribution, with training partitions further partitioned to provide a 15 percent validation monitor. The deep LSTM model comprised an input layer, three progressive hidden layers parameterized with 80, 60, and 32 neurons utilizing rectified linear unit activation, a dropout regularization layer configured at a 0.3 rate to mitigate overfitting, and a terminal five-neuron dense layer activated via softmax. Training was executed across 20 epochs with a batch size of 32 utilizing sparse categorical cross-entropy loss and the Adam optimizer, generating a baseline accuracy of 82.74 percent alongside a cumulative training time of 4,105.5 seconds.

To address the documented class imbalance, subsequent experimental configurations evaluated data augmentation interventions. In the SMOTE trial, synthetic minority sampling expanded all five categorical classes to equal distributions of 93,000 records each, creating an aggregate dataset of 465,000 instances split into 348,750 training and 116,250 testing records; this configuration achieved 80 percent training accuracy and 79.61 percent testing accuracy. In contrast, the GAN configuration utilized a generator incorporating three dense layers of 32, 64, and 128 neurons with LeakyReLU activations and batch normalization, coupled with a discriminator featuring four dense layers of 128, 64, 32, and 16 neurons. GAN augmentation specifically expanded the minority fuzzer class by 10,000 instances to 34,246 records and elevated denial of service records by 15,000 instances to 31,353 records, resulting in 261,995 total entries. When trained on this balanced distribution, the GAN-LSTM model attained a superior overall accuracy of 92 percent, demonstrating that targeted generative synthesis avoids the performance degradation observed during broad linear oversampling while substantially reducing both false positive and false negative classification errors.

Because deep neural architectures inherently operate as opaque predictive systems, the investigators incorporated explainable artificial intelligence frameworks to clarify internal model behaviors and establish verifiable confidence in automated classifications. Local interpretable model-agnostic explanations were deployed to provide instance-level transparency by training localized surrogate models on perturbed feature values around specific predictions, confirming that parameters such as connection state time-to-live and protocol types actively drove correct denial of service classifications. Simultaneously, Shapley additive explanations grounded in cooperative game theory were implemented to compute marginal feature contributions across individual instances and the complete testing distribution. Global SHAP analysis revealed that network connection state, service category, state time-to-live, and destination port counts served as the most influential variables steering final model decisions, with the connection state parameter demonstrating the highest predictive importance. By validating that classifications depended on logically sound flow properties rather than spurious artifacts, the research establishes a reproducible methodology for deploying transparent, high-performing intrusion detection systems across increasingly vulnerable Internet of Things networks.

Subject of Research: Deep learning and generative data augmentation for intrusion detection in Internet of Things networks

Article Title: Enhancing IoT intrusion detection using an LSTM-based deep learning model with GAN and explainable AI

Article References: Sharma, B., Sharma, L., & Lal, C. (2026). Enhancing IoT intrusion detection using an LSTM-based deep learning model with GAN and explainable AI. Discover Artificial Intelligence, 6(1), Article 1867. https://doi.org/10.1007/s44163-026-02657-3

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02657-3

Keywords: Intrusion Detection System, Internet of Things Security, Long Short-Term Memory, Generative Adversarial Network, SMOTE, Explainable AI, SHAP, LIME, UNSW-NB15 Dataset, Class Imbalance, Pearson Correlation, Feature Reduction, Denial of Service Attacks, Fuzzer Attacks, Anomaly Detection, Deep Learning, Network Cybersecurity

News Source: Josephine Dean. (November 13, 2026). Generative Adversarial Networks and Deep Learning Bolster Internet of Things Security. Scienmag.

Tags: Anomaly Detectionclass imbalancedeep learningdenial-of-service attacksExplainable AIfeature reductionFuzzer Attacksgenerative adversarial networkInternet of Things securityintrusion detection systemLIMElong short-term memoryNetwork CybersecurityPearson correlationSHAPSMOTEUNSW-NB15 dataset
Share12Tweet7Share2ShareShareShare1

Related Posts

Mass Timber Engineering Faces New Frontiers in Wood Species Selection

Mass Timber Engineering Faces New Frontiers in Wood Species Selection

November 18, 2026
Sensor Fusion Restores Classifier Accuracy Across Differing Power Line Inspection Robots

Sensor Fusion Restores Classifier Accuracy Across Differing Power Line Inspection Robots

November 18, 2026

Perturbed Supra Metric Framework Proves Solutions for Fractional Disease Model

November 18, 2026

Simplifying Vehicle Shapes in Virtual Crash Simulations Underestimates Occupant Injury Risks

November 18, 2026

POPULAR NEWS

  • Posterior Surgical Resection Demonstrates Safety and Efficacy for Presacral Tailgut Cysts

    Posterior Surgical Resection Demonstrates Safety and Efficacy for Presacral Tailgut Cysts

    29 shares
    Share 12 Tweet 7
  • High Early Burden of Post-Stroke Spasticity Persists Despite Structured Follow-Up Care

    29 shares
    Share 12 Tweet 7
  • Dynamic Bone Marrow Immune Remodeling Drives Leukemia Progression in Myelodysplastic Neoplasms

    29 shares
    Share 12 Tweet 7
  • Reduced Infant Opioid Withdrawal Severity Associated With Care Model Expansion

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Dutch Study Reveals Systemic Gaps Hindering Workplace Chemical Health Risk Collaboration

Low Calcium Triggers Destructive Reactive Oxygen Buildup in Chinese Cabbage

Preventive Prostatectomy May Benefit Selected Men at High Risk of Prostate Cancer

Subscribe to Blog via Email

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

Join 88 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

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
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.