Taylor Walrus Framework Achieves 92.85 Percent Attack Detection Accuracy in Sensor Networks

A research study published in Discover Computing by S. Manikandan and Suganthi Santhanam introduces an integrated computational framework that resolves two critical vulnerabilities confronting modern wireless sensor networks: rapid battery exhaustion during multi-hop communication and severe susceptibility to malicious cyber intrusions. Designated as the Taylor Walrus Optimization Algorithm-based Dense Google Forward Harmonic Network, or Taylor WaOA_DGFHNet, the methodology unifies reinforcement learning, bio-inspired metaheuristics, and dual-branch deep neural networks. Experimental evaluations reveal that the unified system achieves an attack classification accuracy of 92.850 percent, a sensitivity of 95.205 percent, a precision of 92.922 percent, and a false positive rate restricted to 7.505 percent when evaluated at ninety percent training data on benchmark network records. Furthermore, during specialized blackhole intrusion simulations, the architecture achieved a targeted detection rate of 93.200 percent with a false negative rate of only 6.800 percent. By pairing predictive energy routing with advanced feature representation at the central base station, the framework provides a mathematically validated solution that protects monitoring environments while preserving operational longevity.

Wireless sensor networks comprise numerous spatially dispersed, autonomous sensor nodes deployed across physical environments to measure dynamic ecological phenomena such as temperature, barometric pressure, acoustic signals, motion, and humidity. These distributed nodes collect situational telemetry and route the observed information through multi-hop wireless relays to a central base station, where the aggregated data supports real-time decision-making across civilian, industrial, and defense infrastructure, including nuclear installations, chemical plants, smart agriculture, transportation grids, and disaster response zones. However, individual sensor nodes operate under severe hardware and environmental constraints, characterized by non-rechargeable battery power, modest memory capacities, limited computational throughput, and vulnerable radio ranges. In standard operational cycles, wireless transmission alone accounts for approximately eighty percent of total node power expenditure, with the remainder consumed by sensing, internal processing, and channel listening. Consequently, unbalanced traffic routing can rapidly exhaust battery supplies, precipitating premature node failures, partitioned communication topologies, and severe operational blind spots across the monitored deployment zone.

Compounding these resource constraints, the arbitrary spatial distribution and open physical accessibility of sensor nodes expose wireless sensor networks to hostile exploitation by external adversaries and compromised internal entities. Malicious intrusions—such as wormholes, Sybil spoofing, distributed denial-of-service incursions, and blackhole attacks that selectively or completely drop routed packets—can corrupt sensor data integrity, subvert forwarding paths, and inject deceptive telemetry that compromises critical administrative decisions. Conventional defense mechanisms historically depended on handcrafted feature engineering, node-behavior scoring, or rudimentary machine learning classifiers that struggle to model complex correlations in high-dimensional traffic streams while burdening low-power nodes with excessive computation. To overcome these deficiencies, the Taylor WaOA_DGFHNet system establishes a distinct division of operational labor. Lightweight sensor nodes focus strictly on physical telemetry sensing and structured forwarding, while resource-intensive analytical tasks—including deep reinforcement learning for energy forecasting, geometric feature fusion, data bootstrapping, and dual-branch neural classification—are completely centralized at the base station, which possesses the requisite computational power and memory capacity.

The foundational component of the communication layer is an advanced routing protocol governed by the Taylor Walrus Optimization Algorithm. Classical Walrus Optimization simulates the natural social behaviors, foraging migrations, feeding routines, and territorial combat observed in Arctic walrus populations when melting sea ice drives mass movements toward coastal outcroppings. In the authors’ formulation, the exploratory phase of this population-based metaheuristic is augmented by incorporating mathematical Taylor series expansions. This enhancement refines candidate position updates across iteration steps, accelerating algorithmic convergence while lowering the numerical overhead required to identify optimal transmission paths. During route selection across candidate node arrays, the routing engine evaluates a multi-parameter objective fitness function that scores potential multi-hop configurations. Rather than relying solely on static residual battery measurements, the fitness metric integrates five distinct operational variables: inter-node transmission distance, channel throughput, node trustworthiness, current energy expenditure, and predictive energy consumption estimates generated through deep reinforcement learning, ensuring that chosen paths minimize latency while avoiding prematurely depleted nodes.

Energy prediction within the routing routine is executed by a Deep Q-Network, or DQN, which models the impending power dissipation of candidate sensor nodes. The predictive architecture incorporates convolutional, dense, flattening, and dropout layers trained via experience replay and temporal-difference learning, calculating expected energy depletion through Q-value estimation to prevent policy divergence during reinforcement cycles. By deriving a predicted residual energy metric defined as one minus consumed energy, the DQN provides future-oriented awareness that allows the Taylor WaOA routing mechanism to circumvent nodes approaching exhaustion before routing failures occur. Operating in parallel with energy modeling is a dual-tier trust evaluation mechanism designed to isolate malicious relays. The framework computes direct trust from empirical interaction records, weighing parameters such as unselfish node conduct, topological rank, residual energy, and packet forwarding depletion. When direct neighbor interactions are sparse, indirect trust is evaluated through shared mutual intermediate nodes, consolidating neighborhood recommendations to formulate an objective composite trustworthiness score that steers multi-hop paths away from compromised entities.

Once network traffic and routing logs arrive at the base station, the data stream enters an intensive preprocessing sequence designed to eliminate statistical noise and diminish dimensional complexity prior to deep classification. High-dimensional sensor records frequently contain redundant metrics that distract deep neural networks and degrade classification efficiency. To resolve this challenge, the framework executes feature fusion utilizing a lightweight LeNet convolutional neural network integrated with geometric Chord distance measurements. Candidate and target features extracted from network records are systematically sorted and evaluated using Chord distance formulas to assess geometric feature similarity and purge redundant operational parameters. The filtered feature representations are subsequently processed through the convolutional, pooling, dense, and flattening layers of LeNet to determine adaptive fusion weighting coefficients. By compressing input feature matrices from an initial dimensional scale into a condensed, discriminative representation, this fusion stage preserves vital spatial and behavioral patterns while substantially diminishing the computational overhead imposed on subsequent deep learning stages.

Following feature fusion, the condensed network records undergo synthetic data augmentation through a statistical bootstrapping methodology. Conventional data augmentation techniques developed for computer vision—such as image flipping, cropping, spatial rotation, and horizontal shearing—are fundamentally incompatible with structured numerical network traffic records that represent numerical flow metrics, byte tallies, protocol types, and packet timestamps. The bootstrapping procedure addresses this operational constraint by applying non-parametric resampling with replacement directly to the fused numerical records. This process expands sample volume while rigorously preserving the intrinsic statistical distributions, inter-feature correlations, and underlying variances of the original dataset. Crucially, the bootstrapping mechanism is utilized exclusively on the training data partitions to rectify severe class imbalances by expanding the representation of rare, underrepresented intrusion types, such as stealthy routing attacks. By contrast, validation and experimental test sets are strictly maintained in their original, unmodified forms to guarantee unbiased, objective empirical assessments of real-world generalization.

At the core of the attack detection pipeline is the newly designed Dense Google Forward Harmonic Network, or DGFHNet. The hybrid deep learning architecture establishes two parallel processing branches to simultaneously exploit complementary feature extraction paradigms: a DenseNet branch and a GoogLeNet branch. The DenseNet pathway employs dense layer connectivity, directly coupling every convolutional layer to every subsequent layer within dense blocks. This architectural layout maximizes gradient flow, prevents vanishing gradient degradation during backward propagation, and facilitates extensive feature reuse across intermediate representations. In parallel, the GoogLeNet pathway processes augmented network records across nine specialized Inception modules spanning twenty-two distinct layers, deploying multi-scale parallel convolutional kernels to capture intricate behavioral patterns across varying operational granularities. The parameters and hyperparameters of the integrated DGFHNet classifier are dynamically optimized during training utilizing the Taylor Walrus Optimization Algorithm to minimize mean squared classification error and achieve rapid convergence.

Rather than combining the outputs of the DenseNet and GoogLeNet branches through conventional feature concatenation or basic linear addition, DGFHNet introduces a novel fusion mechanism based on numerical forward harmonic analysis. Forward harmonic evaluation models complex mathematical relationships and intermittent data variations by decomposing intermediate feature vectors into structured trigonometric components based on cosine and sine functions. By processing the intermediate representations through harmonic series expansions, the network mathematically derives harmonic coefficients that capture subtle shared and complementary dependencies between dense spatial features and multi-scale Inception representations. The resulting forward harmonic synthesis yields a consolidated output representation that captures non-linear interactions between disparate network characteristics far more effectively than standard concatenation schemes. This unified harmonic representation is ultimately projected through dense output layers to generate definitive classifications, successfully distinguishing benign transmission patterns from complex, multifaceted cyber intrusions with high discriminatory precision.

The research team validated the proposed framework through extensive experimental simulations implemented in Python across three benchmark intrusion datasets: BoT-IoT, NSL-KDD, and WSN-DS. In 2000-round routing trials on the BoT-IoT environment, Taylor WaOA achieved a residual energy level of 0.2350 Joules, an average transmission distance of 35.48 meters, a network trust score of 0.958, and a throughput of 0.993 Megabits per second, substantially surpassing baseline algorithms such as CSDP, Taylor C-SSA, SARP, LEACH-ISMO, SFO-SHO, EMRAR, and EMO-PEGASIS. In attack classification with ninety percent training data on NSL-KDD, Taylor WaOA_DGFHNet achieved an accuracy of 92.850 percent, sensitivity of 95.205 percent, specificity of 92.495 percent, precision of 92.922 percent, an F1-score of 94.050 percent, and a Matthews Correlation Coefficient of 0.918, while restricting the false positive rate to 7.505 percent. Comparable accuracy was recorded on WSN-DS at 92.238 percent and BoT-IoT at 91.708 percent, with receiver operating characteristic evaluations reaching area-under-the-curve true positive rates of 0.999 to 1.000. Processing overhead remained modest, requiring execution times between 1.48 and 2.01 seconds and memory allocations between 116 and 137 Megabytes, with statistical validity confirmed via ANOVA and paired t-tests at p-values below 0.05. Despite these achievements, the authors highlight operational limitations, including base station single-point failure vulnerability, omission of harsh environmental modeling, and reliance on synthetic IT traffic rather than physical testbeds.

Subject of Research: Intrusion detection and energy-efficient routing in wireless sensor networks using hybrid optimization and deep learning

Article Title: Attack detection in WSN using Taylor Walrus Optimization Algorithm-based Dense Google Forward Harmonic Net

Article References: Manikandan, S., & Santhanam, S. (2026). Attack detection in WSN using Taylor Walrus Optimization Algorithm-based Dense Google Forward Harmonic Net. Discover Computing, 29(1), Article 871. https://doi.org/10.1007/s10791-026-10755-0

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10755-0

Keywords: wireless sensor networks, Taylor Walrus Optimization Algorithm, Dense Google Forward Harmonic Network, intrusion detection, energy-efficient routing, Deep Q-Network, LeNet feature fusion, Chord distance, bootstrapping data augmentation, DenseNet, GoogLeNet, forward harmonic analysis, BoT-IoT dataset, NSL-KDD dataset, WSN-DS dataset, blackhole attack, trust model, cybersecurity