Conferences Links
|
|
Paper Review Claims |
| # |
Paper ID |
Title |
Abstract |
Congress |
Claim |
| 1 |
79.Cnf-1253 |
Electric Vehicle Routing with Time-Dependent Waiting Time at Charging Stations Using Stochastic Optimization |
This paper addresses electric vehicle routing problems with deterministic and stochastic waiting times at charging stations. A new model is proposed to account for time-dependent queue delays to minimize total operating costs and penalties for late deliveries by optimizing routes and charging-station delays. The solution was proposed using a two-stage stochastic optimization methodology that combines adaptive large neighborhood search with a simulation-based approach to handle the complexity of the problem. Experimental results show that considering time-dependent stochastic waiting times in fleet planning results in more robust plans, lower total cost, and better customer experience due to fewer late deliveries. Real-time or dynamic routing could be a future extension in addition to non-homogeneous fleet considerations. |
EngiTek 2026 |
Claim |
| 2 |
80.Cnf-1255 |
Design and Implementation of a Web-Based Student
Management System for the Faculty of Engineering,
University of Abuja: An Agile (XP-Informed) Approach |
The Faculty of Engineering, University of Abuja,
serves over 1,240 undergraduates across eight departments, yet
course registration, result checking, transcript issuance, and
Students Industrial Work Experience Scheme (SIWES)
monitoring still rely on manual, fragmented processes. This paper
presents the design and implementation of a web-based Student
Management System that consolidates these functions into one
role-based platform for administrators, students, and alumni.
Development followed an Agile, Extreme Programming (XP)-
informed approach over five sprints in twelve weeks, using a
Model-View-Controller (MVC) architecture on PHP 8, MySQL
8.0, and Bootstrap 5.3, with the FPDF library generating
registration slips, result statements, and transcripts. The system
delivers role-based access control (RBAC), AJAX-driven course
registration with real-time unit-load validation, GPA/CGPA
computation, a SIWES logbook and evaluation module, and a
transcript request-and-tracking workflow. Testing recorded an
average page load time of 1.2 s, database query times under 100
ms, PDF generation in 0.8 s, a 98% course-registration completion
rate, and zero critical vulnerabilities in OWASP ZAP penetration
tests. The results show that an integrated, Agile-developed
platform can reduce administrative overhead and extend roundthe-clock access to academic services in similarly resourced
Nigerian university faculties.
|
EngiTek 2026 |
Claim |
| 3 |
55.Cnf-1213 |
Parametrically Efficient Conv‑TasNet for 44.1 kHz Speech Separation |
Speech separation is a fundamental task in audio processing, yet most state-of-the-art models are trained on low-frequency datasets (8–16
kHz) and are not readily adaptable to high-quality
audio (44.1 kHz). In this work, we propose
a modified Conv-TasNet architecture adapted
for high-fidelity speech separation at 44.1 kHz.
Our contributions include architectural modifications — replacing global layer normalization with
GroupNorm, introducing dynamic padding for
variable-length inputs, and systematically tuning
hyperparameters (kernel size, stride, TCN depth).
We generated a custom VCTK-2mix dataset comprising 30,000 training mixtures with controlled
SNR and implemented a fully reproducible pipeline
in PyTorch. Extensive experiments demonstrate
that our model achieves an SI-SNR of 16.88 dB,
SDR of 17.33 dB, SIR of 26.38 dB, and SAR of 17.92
dB, surpassing the original Conv-TasNet (15.3 dB
on WSJ0-2mix at 8 kHz). Our findings indicate
that increasing kernel size relative to the original
architecture is beneficial for 44.1 kHz processing,
while maintaining a compact model size suitable
for resource-constrained environments. The source
code is publicly available to ensure reproducibility. |
EngiTek 2026 |
Claim |
| 4 |
55.Cnf-1214 |
Application Of Rüppell's Fox Optimizer for Solving Economic Dispatch in Power System |
Economic Dispatch (ED) allocates power generation among thermal units to minimize total fuel cost subject to operational constraints. The resulting optimization is made non-linear and non-convex by real-world factors such as valve-point loading effects, transmission losses, and emission penalties. In this paper, the Rüppell's Fox Optimizer (RFO), a metaheuristic inspired by the foraging behaviour of Rüppell's foxes, is applied to constrained ED problems. RFO uses a multi-sensory search mechanism that combines eyesight, hearing, and scent-driven strategies to transition between exploration and exploitation. Two standard test systems are used for validation: a 6-unit IEEE 30-bus system with quadratic cost functions and B-coefficient transmission losses, and a 10-unit system with valve-point effects, losses, and emission penalty cast as a single-objective Combined Economic Emission Dispatch (CEED) problem. On the 6-unit system, a best fuel cost of 800.9034 $/hr was achieved by RFO, compared to PSO (806.1893 $/hr) and WOA (809.3234 $/hr), with statistical significance confirmed by the Wilcoxon rank-sum test (p < 0.001). On the 10-unit CEED, the lowest fuel cost of 112,886.6560 $/hr was recorded by RFO. Over 30 independent runs, RFO is shown to be a competitive solver for ED problems in power systems. |
EngiTek 2026 |
Claim |
| 5 |
55.Cnf-1216 |
Wavelet-Based Multiresolution Analysis for Fault Detection in Nonlinear PMSM Drives A Finite Element Study |
Wavelet-based multiresolution analysis is utilized
for fault diagnosis in permanent magnet synchronous motors
operating under direct torque control. The inherent challenges of this control method, namely variable switching frequency and hysteresis tolerance, are addressed by applying the maximum overlap discrete wavelet transform to the estimated electromagnetic torque signal. This noninvasive approach extracts time-frequency characteristics of the torque signal while isolating
the switching and hysteresis impact in separate bands to ensure clear identification of fault related components. The RMS values of the wavelet detail coefficients serve as inputs to a quadratic discriminant analysis classifier that not only detects and differentiates common faults including turn-turn short circuits, mixed eccentricity, and partial demagnetization but also estimates their severity levels. Validation is performed through high fidelity simulations using ANSYS-MAXWELL involving a finite element machine model integrated with the drive system implemented in ANSYS-SIMPLORER, under both normal and faulty conditions. This method also supports the creation of fault detection databases in practical settings, mitigating the scarcity
of experimental fault data. |
EngiTek 2026 |
Claim |
| 6 |
80.Cnf-1220 |
HAI-KDG: A Computable Framework for Measuring the Human–AI Knowledge Divergence |
Abstract— Generative artificial-intelligence (AI) systems now produce candidate knowledge faster than human experts can audit it, creating a widening and largely unmeasured asymmetry between what machines produce and what humans can comprehend, verify, and use. We argue that this asymmetry should be measured rather than described. Therefore, we introduce HAI-KDG, a framework which formalized as a five-component vector: scale, novelty, assimilation cost, verifiability, and rate of change. In this paper, three components of the framework are computed and validated, while the remaining two are formally specified with a calibration protocol fixed in advance. An explicit proximity criterion defines the machine-unique set, a two-level novelty measure combines distributional divergence with a cross-corpus neighbor-distance ratio (CNDR), and a five-layer architecture keeps its upper layers portable across domains. Directed single-stimulus tests on synthetic corpora with known ground truth show, over ten seeds, that each computed dimension responds selectively to its own stimulus and is robust to the framework's free parameters. A complete 3×3 grid of real-corpus studies exposes a dissociation that replicates across every condition: at lexical embeddings, fluent machine text diverges distributionally in seven of nine conditions (up to 9.2 control standard deviations, Holm-adjusted p < 0.05) while instance-level isolation stays near chance, a pattern that survives a change of lexical representation. Generated code inverts it, separating at both levels and reaching 0.986 ± 0.007 AUROC, where CNDR leads the density-based baselines but is indistinguishable from nearest-neighbor distance; its clear advantage over all baselines appears instead in the weak-signal text conditions. |
EngiTek 2026 |
Claim |
| 7 |
55.Cnf-1222 |
A Compact Tri-Band Printed Antenna for Future Wireless and 6G-Oriented Applications |
This paper presents a compact tri-band printed antenna based on a branched planar radiator, a tapered microstrip feeding structure, and a Partial Ground Plane. The antenna is designed on a 30 × 30 mm² Rogers RO4003C substrate with a thickness of 0.813 mm, relative permittivity of 3.38, dielectric loss tangent of 0.0027, and 0.035 mm copper thickness. The geometry was developed to produce three resonant operating points around 2.45, 3.5, and 5.8 GHz using different effective current paths in the branched radiator. The tapered feed provides a gradual transition from the wider 1.8 mm input section to the narrower 1.1 mm radiator-side section, reducing the impedance discontinuity and supporting the matching between the excitation line and the radiating structure. The three bands have VSWR of 2.55, 1.64 and 1.46 respectively, and realized gain of 1.12, 3.02 and 3.13 dB accordingly. The total efficiency of 82.1, 82.3 and 90.4 % are recorded at 2.45, 3.5 and 5.8 GHz respectively. Far-field results show frequency-dependent radiation characteristics, while the simulated current distributions demonstrate that field concentration changes with frequency and follows different portions of the branched radiator. The proposed structure provides a simple single-element approach for multi-band wireless. |
EngiTek 2026 |
Claim |
| 8 |
80.Cnf-1221 |
Using Hussein Search Algorithm for Optimizing Search in
Metadata of IDS |
Metadata analysis is the process of examining metadata
to discover patterns, relationships, trends, and useful information
about the underlying data. Query latency is
the total time from when a user submits a query until
the result is returned. It typically includes:Network Delay,
Query Parsing, Index Lookup, Search Time, Ranking,
Result Formatting,Transmission. Therefore, reducing
search time, or index lookup can reduce query Latency.
This paper explains how The Hussein Search Algorithm
is used for efficient search in Intrusion Detection Systems
(IDS) metadata. We evaluate its performance on synthetically
generated uniform and skewed metadata distributions
and compare it against classical algorithms including
binary search and interpolation search. Our results
demonstrate that the Hussein Search Algorithm achieves
expected constant-time average complexity with lower
memory overhead than hash-based methods while maintaining
competitive speed. The performance of Hussein
Search Algorithm (HSA) in term of search time per query
outperforms binary search algorithm, where it is 2.3-2.6
times faster than binary search algorithm for uniform data
distribution and highly faster than interpolation search for
skewed metadata. In addition, HSA maintains its advantage
in search time when the datasets size is increased,
whether, the searching time of the other algorithms increases
with the growth of the metadata size. |
EngiTek 2026 |
Claim |
| 9 |
79.Cnf-1223 |
Enhancement of the structural and optical properties of perovskite solar cells by the implementation of different nanostructured electron transmission layers |
The purpose of such research is to comprehensively
investigate the impact of various structures of Electron
Transmission Layer (ETL) on the optical and structural
properties of Perovskite solar cells. Four structures of electron
transmission layers were investigated with controlled
thicknesses. These structures were a single-layer of ⁓ 50 nm
ZnO, a single-layer of ⁓ 50 nm TiO2, Bi-layer of TiO2 / ZnO
(10/50 nm), and a Tri-layer of TiO2 / ZnO / TiO2 (10/50/10 nm).
UV-Vis spectroscopy revealed that the ZnO thin film of ⁓ 50
nm thickness transmits higher amount of visible light toward
Perovskite thin film than its counterparts, while TiO₂ ETLs is a
perfect UV light-blocker. XRD and SEM analyses showed that
a ZnO single layer enhanced perovskite crystallinity up to ~
89.5% and reduced strain by ~ 96%, making a smoother, more
uniform polycrystalline surface. These findings confirm ZnO as
the most efficient ETL for improving PSC performance, and
TiO₂ offering additional UV protection. This work is a
synergetic marriage between nanotechnology and the perfect
light-absorbing material – Perovskite, –where a very efficient
light-harvesting device with high crystallinity, higher stable and
less amount of defect is produced. |
EngiTek 2026 |
Claim |
| 10 |
79.Cnf-1225 |
Rule-Based Closed-Loop Service-Capacity Control
for Digital Queue Management in Government
Service Centers |
Abstract—Long and time-varying queues remain a practical
problem in government service centers, particularly when service
capacity is fixed while visitor demand changes during the day.
This paper presents a web-oriented digital queue architecture in
which ticket events provide feedback to a rule-based closed-loop
controller. The measured waiting time is compared with a desired
operating range, and the controller recommends increasing or
decreasing the number of active service counters. A time-stepped
evaluation compares a fixed two-counter policy with the proposed
policy over a 120-min demand profile. Under the same demand
and per-counter service-rate assumptions, the proposed controller
reduces the reported average waiting time from 5.91 to 1.69 min
and the maximum queue length from 76.32 to 47.19 visitors.
These changes correspond to reductions of 71.4% and 38.2%,
respectively, while the average number of active counters rises
from 2.00 to 2.25. The study therefore demonstrates a clear
waiting-time–capacity tradeoff and shows how ordinary digital
check-in records can support closed-loop operational decisions
without physical crowd sensors. The results are simulation-based
and are intended as a proof of concept rather than a field-validated
deployment.
Index Terms—digital queue management, feedback control,
government services, rule-based control, service capacity, waiting
time |
EngiTek 2026 |
Claim |
| 11 |
80.Cnf-1227 |
FED-Sentinel: Explainable Federated Multi-Modal
Learning for Evidence-Based Fault Detection in
Kubernetes Microservices |
—Kubernetes-orchestrated microservice deployments
fail in ways that outpace threshold-based monitoring: faults
cascade across loosely coupled, dynamically scheduled services
faster than a human operator, or a static alerting rule, can react.
This paper reports on FED-Sentinel, a multi-modal, federated,
explainable fault-detection pipeline that fuses container metrics,
Kubernetes events, distributed traces, and logs into a single
feature representation, scores that representation with a temporal
anomaly stage feeding an XGBoost–LightGBM stacking ensem
ble, aggregates model updates across simulated clusters under
differential privacy, and attaches SHAP-based explanations to
every prediction. Unlike prior work in this space, we separate
every reported number into two explicit classes: design targets
quoted against nine literature systems, and an independently
executed evaluation– five seeds, paired significance testing– against nine conventionally fitted detectors under identical
conditions. The executed pipeline reaches F1 = 0.9883 but is
not statistically distinguishable from a well-tuned Random Forest
(p = 0.11). An executed ablation shows the temporal anomaly
stage is not accuracy-additive and is the single most expensive
stage per inference; cross-workload zero-shot transfer collapses
(F1 = 0.1603); and differentially private federation is usable
only for privacy budgets ε ≥ 4. We report these negative and
boundary results alongside the positive ones and argue that an
honest accounting of which architectural components earn their
computational cost is itself a necessary contribution for systems
intended for production remediation. |
EngiTek 2026 |
Claim |
| 12 |
80.Cnf-1246 |
IntentCheck: A Lightweight Tool for Detecting Intent Violations During Code Review |
Conventional code review support tools, static checklists and rule-based analysers such as Pylint and SonarQube, focus on what the code is rather than what it was intended to be. We present IntentCheck, a lightweight, offline Python tool that dynamically generates change-specific review checklists targeting intent violations: code that is syntactically valid and passes static analysis, yet does not faithfully reflect the developer’s design intent. Rather than filtering whole-file warnings to the changed lines, IntentCheck is diff-driven: it classifies every hunk of a diff, analyses the enclosing scope of each functional change with ten Abstract Syntax Tree (AST) checkers, and follows changed definitions to the files that use them. It requires no AI inference, no cloud connectivity and no third-party dependencies, making it suitable for regulated safety-critical environments. On 100 randomly sampled commits from two open-source projects that were not used during development, it produced about two checklist items per commit; one item in ten, and 23% of non-informational findings, warranted reviewer action. Compared with diff-scoped pyflakes, Ruff and Pylint, 94% of its actionable items were reported by none of them, while each approach caught developer-confirmed defects that the other missed: IntentCheck complements rather than replaces linters. |
EngiTek 2026 |
Claim |
| 13 |
80.Cnf-1245 |
Stress-Testing an Evidence-First Bilingual Hiring Ranker: Keyword Stuffing and the Accuracy Trade-off in Arabic and English |
We study Muwafaq, a research prototype that ranks curriculum vitae (CV) documents for Arabic and English job ads by matching each job requirement to the best supporting work bullet or skills-list item, and that returns those evidence spans so a person can check each match. We stress-test it on a small synthetic benchmark of 250 CV and job pairs (25 jobs) with human grades. Its ranking quality, measured by normalised discounted cumulative gain at rank 10 (nDCG@10), is comparable to BM25 and higher than whole-CV embedding baselines. However, one visible keyword-stuffing line moves weak CVs up by about 4 to 5 places among ten, and every simple fix that cuts this gain also cuts accuracy. A copy-gate plus a gap penalty lowers the worst case over 20 attacks on 13 test jobs to about 2 places. A pool-level check, which compares each CV with the other CVs for the same job, lowers it to 1.1 (E5) and 0.7 (BGE-M3) places, at 0.035 to 0.047 lower nDCG@10 than the undefended score; this is close to, but not fully within, our 1-place and accuracy targets, and fake work bullets remain its worst case. Capacity-style and requirement-anchored fixes did not help, and a language-model judge loses accuracy and accepts fake bullets. We also report location, name and gender-wording checks and a small check on real CVs. We are not aware of a validated defence for visible keyword stuffing in CV ranking; ours is a partial mitigation. |
EngiTek 2026 |
Claim |
| 14 |
81.Cnf-1248 |
Comparative Assessment of Conventional and Microplane-Based Nonlinear Finite Element Modeling of Reinforced Concrete Beams in ANSYS |
Many civil engineering structures are constructed from Reinforced Concrete (RC). RC structures can be subjected to crushing, cracking, and reinforcement yielding, and hence their behavior is nonlinear. As a result, simplified linear analysis is inadequate in many cases. With the development of computers, nonlinear numerical analysis like the Finite Element (FE) Method became widely used. Furthermore, the FE method can be utilized in parametric studies with a lower cost compared to experimental testing. As a result, the FE method is an efficient tool for design decisions. ANSYS is a widely used FE analysis software in many applications. Modeling RC structures in ANSYS is possible using Solid 65 elements as it captures crushing and cracking behavior. However, solid 65 is a legacy element although it is still used in recent research works. In this study, solid 65-based modeling is compared with the new ANSYS technology the microplane model. A simply supported beam is considered with five modeling methods taking into account different rebar formulations. It is shown that the microplane model exhibits a smoother post-cracking response compared to the solid 65-based models. The effect of reinforcement modeling on results is minimal in terms of accuracy, but clear when it comes to numerical efficiency. |
EngiTek 2026 |
Claim |
| 15 |
79.Cnf-1230 |
Volumetric reconstruction of lung cancer infection region based seed-guided fast marching method using computed tomography image processing |
Among all malignancies, lung cancer carries the heaviest mortality burden worldwide, and precise tumor-volume measurement on computed tomography (CT) scans underpins both treatment planning and longitudinal monitoring of the disease. The segmentation approaches dominating the field today are deep-learning-based, yet they demand extensive labeled training sets and offer limited insight into how a given tumor boundary was actually derived. Addressing this gap, the pipeline introduced here is a semi-automatic, seed-guided Fast Marching Method (FMM) that segments the tumor straight from CT images with no training data whatsoever, combining adaptive intensity-based thresholding, geodesic-distance region growing, and morphological refinement, driven by nothing more than a handful of user-selected seed points on each slice. The method was tested on 18 tumor cases from 10 patients, with each automatic result compared against an independently hand-drawn ground-truth mask. The proposed pipeline achieved a mean Dice coefficient of 0.811±0.100, an IoU of 0.692±0.128, a Sensitivity of 0.696±0.132, and a Precision of 0.994±0.010, with Precision exceeding 0.96 in nearly every case. Taken together, these figures demonstrate that a fully interpretable, training-free segmentation method can still deliver strong volumetric accuracy alongside a markedly low false-positive rate, making it a practical alternative to deep-learning-based tumor segmentation wherever annotated data or computing resources fall short. |
EngiTek 2026 |
Claim |
| 16 |
80.Cnf-1232 |
A comparative machine learning approach for predicting cardiovascular disease risk using clinical features |
Cardiovascular disease (CVD) remains a major global health challenge, highlighting the need for accurate and data-driven approaches for risk assessment. This study presents a comparative machine learning framework for predicting a continuous CVD risk score using clinical and demographic features. A dataset of 1,529 samples with 21 candidate features was preprocessed using appropriate imputation, categorical encoding, and feature scaling techniques. To identify the most informative predictors, a consensus-based feature selection framework was applied, resulting in eight key features, including body mass index, systolic blood pressure, total cholesterol, estimated low-density lipoprotein, diabetes status, waist-to-height ratio, fasting blood sugar, and high-density lipoprotein. Eight regression algorithms were evaluated, including Linear Regression, Ridge Regression, Decision Tree, Random Forest, Gradient Boosting, XGBoost, Gaussian Process Regression, and Support Vector Regression. Among the evaluated models, Support Vector Regression achieved the best performance, with a mean absolute error of 0.1328, root mean squared error of approximately 0.290, and an R² of 0.930 on the test set. Cross-validation also demonstrated consistent predictive performance. The findings highlight the potential of machine learning and data-driven feature selection for clinical risk prediction and support the development of reliable predictive tools for cardiovascular health assessment. |
EngiTek 2026 |
Claim |
| 17 |
55.Cnf-1233 |
Design Vision for Developing a Flying X-ray Fluorescence Spectroscopy System |
Traditional methods for geochemical exploration in rough terrain are restricted due to cost, accessibility, and inadequate spatial coverage. This paper presents a conceptual design roadmap for an integrated airborne system using x-ray fluorescence spectroscopy (XRF). This conceptual design depicts the relevance of this technology as an aerial platform capable of conducting in-situ elemental analyses. By creating this vision of how the various components and subsystems (ie. spectrometer, airborne systems, methods of data collection, etc.) could theoretically be integrated into a working airborne imaging hyper-spectral sensor, and creating an example of what a design like this could look like, demonstrates how these systems will operate as one functioning system. The proposed air payload system will also allow geochemists the capability to perform geochemical surveys with increased accuracy and effectiveness in challenging environments. This project is only conceptual and/or exhibits the theoretical limitations, issues, and potential possibilities of this type of system without the use of prototypes or performing real-world experiments. |
EngiTek 2026 |
Claim |
| 18 |
79.Cnf-1234 |
Role of hydrogen in achieving net-zero aviation goals: a review |
The current goals to achieve net-zero aviation face many challenges, including finding alternative fuels with high energy density and meeting strict safety standards. Hydrogen is a promising solution for reducing carbon emissions in aviation. This review discusses the potential of hydrogen to achieve net-zero aviation, focusing on combustion-based propulsion, electric fuel-cell propulsion, and onboard storage technologies. The key advantages and engineering barriers to aircraft integration with hydrogen are compared with jet fuel (kerosene). Combustion-based propulsion has led to innovations in lean-burn combustors, especially for larger aircraft (> 200 seats), while electric propulsion using fuel cells would be a viable option for UAVs and smaller regional aircraft due to its high efficiency and zero combustion emissions. Hybrid propulsion (combustion- and electric-based) is suggested for medium aircraft (~ 150 seats). Despite hydrogen's promise, challenges such as NOx emissions, cryogenic storage, thermal management, and the necessary infrastructure at airports remain to be addressed. Hydrogen may not be the complete short-term solution for low-carbon aviation—like sustainable aviation fuels—but it is expected to be vital in a broader strategy that includes both sustainable aviation fuels and hybrid technologies. |
EngiTek 2026 |
Claim |
| 19 |
79.Cnf-1236 |
Stage-Wise Probabilistic Risk Assessment of a Pressurized Water Reactor Coupled With a High-Temperature Electrolysis Hydrogen Facility |
The integration of light-water reactors with high-temperature steam electrolysis represents a viable approach to producing large commercial quantities of clean hydrogen in the near future, but it is necessary to demonstrate that the heat removal system, the direct electrical connection, and the hydrogen handling plant do not compromise reactor safety before the arrangement can be granted a license. This work carries out a stage-wise probabilistic risk assessment for a typical pressurized water reactor supplying 100, 500 and 1000 MW electrolysis plants, using the general data and fault-tree logic from the Idaho National Laboratory report INL/RPT-23-74319. The coupled plant is divided into three phases and 24 processing stages, the hazards of which correspond to steam-line breaks, transients, loss of off-site power at the switchyard and seismic events. The fault trees for the initiating events have been reconstructed and re-quantified using the smallest possible cut sets, then compared with the acceptance criteria set out in 10 CFR 50.59 and Regulatory Guide 1.174. The re-quantification agrees with three previously published result sets to within 0.3%. The greatest increase in the number of initiators is 7.69% (in the case of a steam-line break with two steam taps), which is below the 10 % limit, and the largest increase in core damage frequency is 1.9E-8 per reactor-year, occurring in Region III. Failure of the normally open isolation valve at each steam tap is responsible for more than 99% of this increase, while hydrogen detonation and generator overcurrent have a negligible effect. Safety margins depend on the number of steam taps, the valve rupture data, and the credited ventilation of the electrolysis enclosures; without that ventilation credit, the 1000 MW facility would increase the loss-of-off-site-power frequency by 10.8%. |
EngiTek 2026 |
Claim |
| 20 |
81.Cnf-1256 |
Urban Open-Space Vulnerability Assessment for Post-Earthquake Evacuation and Emergency Response: A Case Study of Dellys, Northern Algeria |
The 2003 Boumerdes earthquake (Mw 6.8) exposed the consequences of seismic events for both the built environment and the urban systems on which emergency operations depend. This study examines the extent to which roads and non-built urban spaces in Dellys can support population evacuation and emergency assistance after a major earthquake. It applies, to a different urban context, the “Emergency Evacuation and Rescue Facilities” criterion developed within the JICA–CGS seismic micro-zonation study for the Wilaya of Algiers. The assessment is conducted in a geographic information system (GIS), using the framework’s road-accessibility and evacuation-point criteria to examine the spatial relationship between urban areas, passable routes and potential refuge spaces. The resulting maps distinguish areas according to their level of access to evacuation and rescue facilities and reveal where the urban configuration may hinder emergency intervention. The findings provide a basis for identifying locally appropriate measures, including the improvement of route continuity, the provision of accessible refuge areas and the reduction of access constraints in vulnerable neighbourhoods. By focusing on the functionality of open spaces and the road network, the study complements building-centred seismic assessments and contributes to the discussion of urban preparedness in Dellys.
The GIS outputs are used to map the spatial distribution of accessibility constraints affecting evacuation and rescue. They help identify locations where the road network or the distance to potential emergency spaces may limit response operations. On this basis, the study discusses planning measures that could be adapted to local conditions, such as securing alternative routes, improving connections to refuge spaces and addressing bottlenecks that may obstruct emergency vehicles. The study contributes a spatial application of the JICA–CGS emergency evacuation and rescue criterion to Dellys, with particular attention to the relationship between road accessibility and non-built urban areas. It demonstrates the value of considering these urban functions alongside the seismic vulnerability of buildings when preparing local risk-reduction strategies
|
EngiTek 2026 |
Claim |
| 21 |
79.Cnf-1237 |
Structural integrity of spent nuclear fuel cladding under transportation loads: a thematic review and knowledge map |
The cladding of spent nuclear fuel must remain intact during handling and transport for retrieval. This article brings together eight previously published studies concerning the structural integrity of zirconium-alloy spent fuel rods under both normal and hypothetical accident conditions that could occur during transport, grouping the studies into five main themes: degradation-driven property characterization, multiscale structural modeling, uncertain rod-level parameters, failure criteria, and deterministic-probabilistic and margin-based assessments. The theories, methods, findings, and limitations are identified and summarized; a knowledge map is given as both a hierarchical outline and a mind map; and the collection of studies is compared with eleven recent external studies. The resistance of the rods is controlled by pellet-cladding bonding, spacer grid stiffness, and ductility that depends on the fuel cladding characteristics; under strain-based failure criteria, fuel with a high level of burnup has failure probabilities which are many orders of magnitude greater than those of fuel with lower burnup in the case of a 9-m side drop; and the normal transport loads measured are still well below the failure limits. Uniform elongation and hydrogen content are the most closely connected concepts, while the relationships among interfacial bonding, hydride-sensitive failure criteria, microstructural models, and probabilistic assessment remain underexplored. The paper also outlines future research priorities. |
EngiTek 2026 |
Claim |
| 22 |
79.Cnf-1238 |
An early fusion framework for monitoring multi-profile generating processes |
Modern manufacturing processes are monitored by several sensors that produce autocorrelated, multivariate profile signals, which violate the assumptions of classical multivariate control charts. Encoding time series as images and classifying them with Convolutional Neural Networks (CNNs) has proven effective for single-sensor profile monitoring, but its extension to multi-sensor processes remains limited. This paper proposes a multivariate image-based monitoring framework in which each sensor profile is encoded as a Gramian Angular Summation Field (GASF), and the images of all sensors are fused early by stacking them as channels of a single input to a compact convolutional network. The framework is evaluated on a six-sensor semiconductor wafer etching dataset of 1,194 production runs using stratified five-fold cross-validation with three input resolutions of 32, 64, and 100 pixels. The results show that the accuracy and fault detection improved steadily with resolution. At the full resolution of 100 pixels, the early fusion model achieved a mean accuracy of 98.66%, an F1 score of 0.993, and an area under the ROC curve of 0.976, detecting 93% of out-of-control runs while raising false alarms on fewer than 1% of in-control runs, with a training time of about 72 seconds per fold. These results compare favorably with a published Gaussian process benchmark on the same dataset. |
EngiTek 2026 |
Claim |
| 23 |
55.Cnf-1239 |
Adaptive voltage-aware fault injection and assertion-based verification for near-threshold 9T SRAM |
Near-threshold operation significantly cuts down the energy use of SRAM. It also weakens the read and write stability of the memory cells. This makes the bits more prone to faults that depend on voltage levels. Traditional verification methods usually use fixed fault models that do not take voltage into account. This paper introduces an approach—an adaptive voltage-aware fault injection system combined with assertion-based verification—designed specifically for a one-sided Schmitt-trigger 9T SRAM cell.
The supply voltage is divided into four modes: nominal low-voltage, near-threshold and critical. These voltage regions guide an eight-class fault injection model, which runs inside a UVM testbench enhanced with SystemVerilog Assertions (SVA). The study uses a 180-nm CMOS model as its base. Provides estimates based on model assumptions. These include voltage-electrical behaviour, fault injection results, for each voltage region UVM/SVA coverage metrics and Monte Carlo analysis of read static noise margin variations.
The paper compares the method to conventional verification techniques and shows how the new approach better captures real-world voltage-related issues. All results come from simulations and are model-based—no actual silicon data is used. The authors note that validation using Cadence Virtuoso/Spectre and Xcelium is still needed and will be done in the work. |
EngiTek 2026 |
Claim |
| 24 |
55.Cnf-1240 |
Improvement of aliasing rejection in a two-stage decimation filter using certain symmetrical polynomials |
This paper presents a novel two-stage multiplierless decimation filter with improved alias rejection. The first and second stages are corresponding comb filters. Additionally, the second stage cascades two modified comb filters of different lengths. The modified comb filter is obtained by converting to a zero coefficient at the comb's point of symmetry in the second stage. If the comb is odd, it has one point of symmetry; if it is even, it has two. As a result, the overall comb-folding bands gain new zeros, increasing aliasing rejection. We express complexity in the number of adders per output sample (APOS). We illustrate the method with two examples. Comparisons with recent methods demonstrate the benefits of this proposal. |
EngiTek 2026 |
Claim |
| 25 |
90.Cnf-1210 |
Supervised neural network based intrusion detection systems |
The magnificent problem for today’s computer networks is the internet attacks which is growing rapidly. Therefore, applying security methods to prevent such attacks on computer network is very important. Network attacks are challenging because they are continuously changing their patterns and techniques. Machine learning approaches play vital roles in detecting, and preventing attacks in different types of computer networks. The security is a critical issue whereas the security mechanism may affect the way of detecting, analysing, and preventing attacks. Therefore, security methods must modify their techniques to deal with. We have proposed Supervised Neural Network Based Intrusion Detection System (SNN-based IDS) to address a critical and timely issue in computer networking - the escalating threat of internet attacks. As these attacks continue to evolve, finding effective methods to detect and prevent them is of paramount importance. this paper leverages advanced technology to tackle the complex task of intrusion detection. This demonstrates the practical application of cutting-edge methods in cybersecurity. The paper suggests that the proposed model has the potential to predict and mitigate new types of attacks, even those not present in the training dataset. This feature is crucial in a rapidly evolving threat landscape. Our SNN is trained by KDDCUP’99 dataset which include 14 attack types in the test data, with an overall number of 24 training attack types. Experiments results show that the proposed system is capable to identifying the attacks and classifying them with high accuracy and reliability, regardless of the dataset's nonlinearity, size, or incompleteness. |
CompuSys 2027 |
Claim |
| 26 |
79.Cnf-1211 |
On the scalability of PM-EDS maglev systems:from laboratory prototypes to full-scale operation |
As a passive and energy-efficient alternative systemto the maglev technologies; Permanent Magnet ElectrodynamicSuspension (PM-EDS) systems was offered. However, scalabilityremains unclear. This paper present an analytical study of PM-EDS scalability using dimensionless parameters derived fromelectromagnetic relationships. Scaling conditions required topreserve system behavior across different sizes are established,showing that scalability depends on maintaining keydimensionless parameters. The results further reveal that systemperformance is constrained by the coupled effects of velocity,magnetic field strength, and system geometry, as well as regime-dependent behavior. |
EngiTek 2026 |
Claim |
| 27 |
79.Cnf-1212 |
NEUROFLOW: Early ICU delirium prediction via temporal convolutional networks with FHIR-native intervention orchestration |
Intensive care unit (ICU) delirium affects an estimated 7 million United States patients annually and is associated with 30-40% increases in one-year mortality, yet the evidence-based ABCDEF prevention bundle achieves only 0-1% worldwide compliance. Existing machine learning systems generate a risk score and stop, leaving the prediction-to-intervention gap unaddressed. This paper presents NEUROFLOW, a FHIR-native ICU delirium prediction and intervention orchestration system. A 56,065-parameter dilated causal Temporal Convolutional Network (TCN) with a learned pharmacodynamic embedding processes a 40-channel, 63-hour time-series tensor constructed directly from HL7 FHIR R4 resources. Trained on 74,829 MIMIC-IV ICU stays under strict temporal holdout validation, NEUROFLOW achieves AUROC 0.8125 and AUPRC 0.5638, warning 77.4% of eligible patients a median of 20 hours before clinical detection. When predicted risk exceeds a threshold, the system automatically writes a US Core CarePlan with six ABCDEF activities and scores bundle compliance hourly from existing FHIR Observations. Split conformal prediction intervals are encoded as structured FHIR extensions. Zero-shot external validation on 132,900 eICU admissions across 208 community hospitals yielded AUROC 0.6568; a label-noise analysis identified 11,405 clinically plausible undocumented cases, and an in silico sensitivity reclassification to the training prevalence recovered AUROC 0.8893, indicating the external gap reflects documentation incompleteness rather than model failure. |
EngiTek 2026 |
Claim |
| 28 |
80.Cnf-1250 |
Design and Implementation of a Web-Based Student Management System for the Faculty of Engineering, University of Abuja: An Agile (XP-Informed) Approach
|
The Faculty of Engineering, University of Abuja, serves over 1,240 undergraduate students across eight engineering departments, yet its academic administration still relies on manual documentation and fragmented processes for course registration, result checking, transcript issuance, and Students Industrial Work Experience Scheme (SIWES) monitoring. These practices cause delays in transcript processing, errors in course registration, poor accessibility of results, and weak digital oversight of SIWES logbooks and evaluations. This paper presents the design and implementation of a web-based Student Management System (SMS) that consolidates these functions into a single, role-based platform serving administrators, students, and alumni.
Following the Agile principles described by Pressman and Maxim (2014), the system was developed using a sprint-based methodology informed by Extreme Programming (XP) practices, namely small releases, continuous integration, and iterative refactoring, across five sprints over twelve weeks. It implements a Model-View-Controller (MVC) architecture on a PHP 8.x, MySQL 8.0, and Bootstrap 5.3 stack, with the FPDF library generating registration slips, result statements, and transcripts as PDF documents. The modules delivered are multi-role authentication with role-based access control (RBAC), AJAX-driven course registration with real-time unit-load validation, a results module with GPA/CGPA computation and filtering, a SIWES module for logbook submission and supervisor evaluation, and an end-to-end transcript request and tracking workflow.
System testing, comprising unit, integration, user-acceptance, performance, and security testing, recorded an average page load time of 1.2 seconds, database query response times under 100 ms, a PDF generation time of 0.8 seconds, and a course registration completion rate of 98%. The results show that an integrated, Agile-developed web platform can reduce administrative overhead, improve data accuracy |
EngiTek 2026 |
Claim |
| 29 |
79.Cnf-1263 |
Multi-View Threat-Intelligence Fusion Using Deep Semantic Representation Learning |
Cyber threat intelligence frequently appears in multiple heterogeneous forms, including network-flow logs, vulnerability feeds, malware indicators, phishing reports, endpoint telemetry, and analyst-enriched incident descriptions. Nevertheless, conventional security analytics pipelines usually regard these sources as independent units, thereby limiting contextual reasoning and reducing the efficacy of early detection of attacks. This article proposes a multi-view threat-intelligence fusion framework grounded in deep semantic representation learning. The proposed model converts structured flow attributes, temporal activity patterns, protocol-service indicators, directionality measures, and statistical anomaly features into a shared latent representation space. The paper utilizes the Kaggle-available CICIDS2017 machine-learning dataset and tests a semantic fusion neural model alongside Random Forest, XGBoost, and multilayer perceptron baselines. The empirical results indicate that the suggested model was capable of reaching 98.74% accuracy, 98.69% precision, 98.58% recall, and 98.63% F1-score levels, thus exceeding the most powerful classical baseline by 0.91 percentage points in F1-score. In addition, the explainability analysis reveals that among the threat indicators, the ones that weighed the most for decision-making were packet-rate behavior, flow duration, inter-arrival timing, and destination-port patterns. The results show that semantic fusion enhances discrimination between benign and malicious network traffic while preserving interpretability for security operations staff. In this experiment, multi-view fusion is represented through semantic feature groups extracted from CICIDS2017 network-flow data, while future work should validate the framework using heterogeneous SOC sources such as DNS logs, endpoint alerts, vulnerability feeds, malware reports, and natural-language threat intelligence. |
EngiTek 2026 |
Claim |
| 30 |
79.Cnf-1259 |
Hybrid Deep Learning for Flow-Based Intrusion Detection with Applicability to Encrypted Network Traffic |
Encrypted network traffic presents a major challenge for intru-sion detection because payload inspection alone is unavailable when content is protected. This paper presents a hybrid deep learning method for flow-based intrusion detection with ap-plicability to encrypted network traffic. The method uses flow-level statistical features, ordered feature representations, and explainable decision analysis. The proposed architecture uses one-dimensional convolutional layers for local feature extrac-tion, combined with bidirectional recurrent layers to identify sequence-sensitive patterns in the serialized feature representa-tion. Prior to model training, feature normalization and class balancing are performed to reduce bias caused by class imbal-ance. The research uses the publicly available KDD Cup 1999 benchmark as a proof-of-concept proxy and evaluates binary classification performance using accuracy, precision, recall, F1-score, false positive rate, and inference latency. Because KDD Cup 1999 is not a contemporary encrypted-traffic dataset and includes content-derived attributes, the results should not be interpreted as direct validation on modern encrypted traffic. The findings show that the hybrid approach achieved strong performance relative to the evaluated traditional supervised classifiers and single deep learning models in the present exper-imental run. Moreover, the explainability component highlights traffic attributes that most strongly influence detection deci-sions. This paper therefore provides an interpretable flow-based detection workflow whose applicability to encrypted-network scenarios should be validated further on contempo-rary non-payload traffic datasets. |
EngiTek 2026 |
Claim |
| 31 |
79.Cnf-1260 |
Explainable Spam Email Recognition Using Ensemble Learning and Semantic Features |
— Spam email remains a substantial cybersecurity issue as malicious campaigns change vocabulary, delivery timing, ob-fuscation patterns, and persuasive content structures over time. This paper proposes an explainable spam email detection sys-tem that combines semantic feature abstraction, ensemble learning, probability-based risk scoring, and local interpretabil-ity. The intended solution is designed not only to separate spam from legitimate emails but also to explain the reasons for mark-ing a message as suspicious. Since it offers a combination of structured indicators extracted from the content of the email, frequency of characters, and attributes related to the sequences of capital letters which are suitable for reproducible machine-learning evaluation, the authors selected the UCI Spambase dataset for the study. Different algorithms such as logistic re-gression, decision tree, random forest, gradient boosting, sup-port vector machine, and a soft-voting ensemble were used for a comparative experiment. The ensemble model strikes the best compromise among accuracy, precision, recall, F1-score, false-positive control, and receiver operating characteristic perfor-mance. The findings indicate that integrating different learners not only enhances the system's robust handling of ambiguous cases, for instance, when legitimate emails contain commercial terms or spam emails use neutral wording, but also increases overall robustness. Finally, the paper presents an explainable decision layer that transforms model output into evidence un-derstandable to the end-user for security operations. |
EngiTek 2026 |
Claim |
| 32 |
79.Cnf-1261 |
Real-Time Botnet Detection Using Adaptive Feature Engineering and Machine Learning |
Real-time botnet detection is a key requirement for modern cyber-resilient infrastructures since botnets can rapidly alter their communication patterns and blend with legitimate network traffic, whereby they can cause large-scale disruption in cloud, Internet of Things, vehicular, and enterprise scenarios. This article proposes an adaptive feature-engineering and machine-learning framework for detecting botnet activity from network-flow data, with emphasis on low-latency detection, feature relevance, and operational deployability. The proposed system integrates flow normalization, rolling statistical profiling, mutual-information-based feature weighting, drift-aware feature refresh, and ensemble machine learning classification. For empirical validation, the research uses a Kaggle-available botnet traffic dataset structured around benign and malicious network-flow records. Experimental results show that adaptive feature engineering increases the detection capability of classical machine learning models especially when traffic imbalance and changing flow behavior are taken into account. Gradient boosting achieved an accuracy of 0.994, precision of 0.992, recall of 0.991, F1-score of 0.991, and a false-positive rate of 0.008, with an average inference latency of 0.38 ms per flow. The findings imply that adaptive feature engineering continues to be very beneficial as feature quality strongly influences detection speed, false alarm control, and model robustness. |
EngiTek 2026 |
Claim |
| 33 |
79.Cnf-1262 |
AI-Based Cyber Resilience Maturity Assessment for Small and Medium Enterprises |
Small and medium enterprises increasingly rely on digital platforms, cloud services, email systems, mobile communication, and outsourced information technology operations, but many lack formal mechanisms for assessing cybersecurity maturity. This article presents an AI-based cyber resilience maturity assessment framework for small and medium enterprises. The framework integrates governance, operational security, incident readiness, optimization, and machine-learning-based threat detection in a single maturity scoring model. The proposed model measures resilience across five maturity dimensions: data preparedness, detection effectiveness, false positive control, response readiness, and recovery capability. To support the empirical evaluation, the paper uses the widely available UCI Spambase dataset as an example cyber threat-classification benchmark. The experimental findings reveal that tree-based ensemble learning gave the best results, with Random Forest achieving 95.33% accuracy and 94.08% F1-score. The results indicate that artificial intelligence can aid maturity assessment not only by generating prediction scores but also by transforming technical performance into actionable maturity evidence for SME managers. The experiment validates the email threat-filtering component of the proposed maturity framework, while full cyber-resilience maturity validation requires future testing with real SME security, governance, response, and recovery data. |
EngiTek 2026 |
Claim |
| 34 |
79.Cnf-1307 |
Secure Digital Twin Architecture for Edge Enabled Smart Infrastructure |
Edge-enabled smart infrastructure is increasingly relying on cyberphysical sensing, local analytics, distributed energy con-trol, and realtime operational decision-making. Nevertheless, the traditional monitoring architectures often consider digi-taltwin modeling, edge security, data trust, and infrastructure resilience only as separate engineering issues. This study pro-poses a secure digital twin architecture for edge-enabled smart infrastructure in which physical assets, edge devices, risk-scoring services, and twin synchronization layers operate as components of an integrated security framework. The pro-posed architecture enhances the security of smart-building and infrastructure services by integrating device authentica-tion, telemetry validation, lightweight anomaly detection, risk-aware synchronization, and resilience feedback mechanisms. The study uses the UCI Appliances Energy Prediction dataset to assess the possibility of energy-state prediction and anoma-ly-aware twin updating. Findings indicate that ensemble-based edge prediction not only greatly reduces the error as com-pared to baseline regression but also enhances operational stability in the face of sensor drift and replay attack simula-tions. The resulting framework integrates edge intelligence, digital-twin integrity, and smart-infrastructure governance within a unified security architecture. |
EngiTek 2026 |
Claim |
| 35 |
79.Cnf-1267 |
Trust Management Framework for Blockchain-Enabled Supply Chain Cyber Resilience |
Blockchain can enhance supply-chain transparency, non-repudiation, and traceability. However, this does not make supply chains immune from cyber threats. Supply chains can still be affected by various issues such as compromised par-ticipants, misleading transaction records, weak identity as-surance, data poisoning, delayed anomaly discovery, and governance gaps between physical logistics and digital ledg-ers. This article proposes a trust management framework for the cyber resilience of blockchain-enabled supply chains. It integrates a consistent five-layer architecture comprising identity assurance, transaction validation, blockchain com-mitment, adaptive trust scoring, and resilience response with continuous anomaly feedback. Validation of the framework is achieved by using the UCI Online Retail II transactional da-taset, which consists of two years of real online retail transac-tion records from a UK-based non-store retailer, and it is also available on Kaggle. From the dataset transaction features such as transaction, participant, cancellation, product-flow, and monetary-risk were constructed for classifying cyber-resilience-oriented trust. The dataset is used as a transaction-al proxy for validating trust-scoring behavior rather than as a complete blockchain cyberattack dataset. The experimental findings demonstrate that the trust-aware ensemble approach proposed here is more effective in detecting suspicious supply chain transactions than the baseline models logistic regres-sion, decision tree, and the standard random forest. This in-tegration enables supply-chain systems to move beyond static ledger recording toward active resilience, early risk isolation, and evidence-driven recovery. |
EngiTek 2026 |
Claim |
| 36 |
79.Cnf-1275 |
Hybrid Autoencoder and Gradient Boosting Model for Anomaly-Based Security Monitoring |
Anomaly-based security monitoring is required to detect novel attacks for which there are no fixed signatures. However, many operating networks still face issues of high false positives, weak feature representation, and poor adaptability to heterogeneous network behavior. This article proposes a hybrid autoencoder-gradient boosting model for anomaly-based security monitoring in enterprise networks. The autoencoder learns the compressed behavioral representations from the normalized network-flow features, whereas the reconstruction error and the latent em-beddings are passed to the gradient boosting classifier to identi-fy normal and malicious activities. This method aims to harness the representational power of deep learning together with the tabular decision boundary strength of the ensemble learning. Experiments were conducted using the KDD Cup 1999 10% corrected labeled connection-record subset containing 494,021 network connections. The results obtained indicate that the pro-posed hybrid model enhances the detection accuracy, F1-score, and false-alarm control in comparison to logistic regression, random forest, gradient boosting, and a standalone autoencod-er-thresholding baseline. The proposed method achieved 99.14% accuracy, 98.91% precision, 99.26% recall, and 99.08% F1-score in the binary anomaly-detection setting. Such results imply that the hybrid latent-representation learning method is capable of enhancing the practical security monitor-ing situations where both the sensitivity to unknown attacks and the reduction of the analyst's workload are necessary. |
EngiTek 2026 |
Claim |
| 37 |
79.Cnf-1268 |
Lightweight Intrusion Detection for Resource-Constrained IoT Devices Using TinyML |
Resource-constrained Internet of Things devices require intru-sion detection capabilities that can work close to the sensors, gateways, and embedded controllers without relying complete-ly on cloud inspection. This paper introduces a TinyML intru-sion detection framework that is very lightweight and suitable for constrained IoT environments where memory, latency, energy, and connectivity are very limited. The suggested method technically involves compact feature selection, edge normalization, small-footprint classification, threshold-aware alerting, and model refreshing with the help of a gateway. TinyML can be considered a practical and effective comple-ment to security analytics at cloud-scale since it limits the number of exposed packets, enhances local response time, and allows for continuous monitoring even when network connec-tions are disrupted. The experimental part of the paper uses the UCI Breast Cancer Wisconsin Diagnostic dataset as a bi-nary structured benchmark and interprets it as an edge anom-aly-detection proxy to examine Gaussian Naive Bayes, int8-oriented logistic regression, shallow decision tree, and compact random forest models under resource-oriented constraints. The int8-oriented logistic regression model was the most bal-anced, reaching 96.50% accuracy, 95.05% F1-score, and 98.34% AUC while using 31 learned parameters. The bench-mark is used only to assess the computational feasibility of compact binary classifiers; because it does not contain net-work-attack traffic, the reported results should not be inter-preted as direct evidence of IoT intrusion-detection effective-ness. The results therefore suggest the potential suitability of compact classifiers for subsequent evaluation on embedded-class IoT hardware. |
EngiTek 2026 |
Claim |
| 38 |
79.Cnf-1295 |
Lightweight Malware Classification Using Feature Selection and Optimized Neural Networks |
Lightweight malware classification is increasingly necessary for various devices such as edge computing devices, endpoint agents, Internet of Things gateways, and other limited resource security appliances where using traditional deep malware analysis methods may be prohibitively expensive. The article introduces a lightweight malware classification system that integrates feature selection, neural network optimization, and efficient inference to identify malware in binaries. The framework is based on static and dynamic behavioral features from the UCI Malware Static and Dynamic Features VxHeaven and VirusTotal dataset and reduces feature dimensionality prior to training an optimized multilayer neural classifier. The paper mainly addresses the challenge of maintaining a high detection accuracy while reducing the computational work, memory usage, and false alarm rates in real-world cyber-defense scenarios. The approach comprises several stages including data preprocessing, variance filtering, mutual-information ranking, recursive feature elimination, neural architecture tuning, and threshold calibration. The optimized lightweight neural network demonstrated excellent classification capability despite using a smaller feature set, thereby confirming that even small feature representations can be sufficient for malware detection and at the same time do not require very large models. The resulting framework offers a deployment-oriented malware-classification approach for resource-constrained environments in which accuracy, inference efficiency, interpretability, and robustness must be balanced. |
EngiTek 2026 |
Claim |
| 39 |
79.Cnf-1272 |
Generative-AI-Assisted Threat Hunting with Human-in-the-Loop Validation |
Generative artificial intelligence is increasingly being explored as a support layer for cyber threat hunting because it can summarize heterogeneous alerts, interpret log sequences, generate investigative hypotheses, and assist analysts in triaging complex security events. Nevertheless, fully autonomous generative AI carries risks in security operations because hallucinated explanations, incomplete contextual reasoning, and weak validation can mislead analysts during time-critical investigations. This paper proposes a generative-AI-assisted threat-hunting framework that incorporates human-in-the-loop validation as a compulsory control layer between AI-generated hypotheses and operational security decisions. The framework includes log preprocessing, behavioral feature extraction, machine-learning detection, generative reasoning, analyst feedback, and validated escalation. The experiment validates the machine-learning detection layer of the proposed framework, while the generative explanation and human-validation components are presented as a framework-level design requiring future evaluation in operational security operations center (SOC) environments. Using the KDD Cup 1999 benchmark subset, Random Forest achieved 99.31% accuracy and a 99.18% F1-score. This work proposes that generative AI can support threat hunting when its outputs are constrained by structured evidence, transparent reasoning, and human validation. |
EngiTek 2026 |
Claim |
| 40 |
79.Cnf-1269 |
Optimized Feature Selection for Low False Positive Intrusion Detection Systems |
Intrusion detection systems are designed and deployed with the goal of accurately detecting harmful network activities while not creating a flood of false alarms that security operators cannot handle. The feature selection process is a key element of devel-oping such intrusion detection systems to deliver a low false-positive rate. This article presents an optimized feature-selection framework that includes the utilization of the publicly available UCI dataset and comparative evaluation using different ma-chine learning techniques. The framework consists of several components: preprocessing, categorical encoding, variance filtering, mutual information ranking, recursive feature elimina-tion, and classifier-level validation. This leads to the identifica-tion of a compact feature set that preserves attack-detection capability while substantially reducing the misclassification of non-attack flows. The KDD Cup 1999 intrusion detection da-taset available at UCI has been chosen as the experimental dataset because it is composed of network connection records with very fine-grained labeling of normal and attack classes which makes it very suitable for controlled IDS experimenta-tion. The experiments show that an operational IDS perfor-mance, particularly the false positive rate, can be improved through optimized feature selection, which also helps the IDS keep a good level of accuracy, precision, recall, and F1-score, even though it is trained on far fewer features as compared to those in the full-feature setting. The highest-performing opti-mized setting gave 99.31% accuracy, 99.46% precision, 99.57% recall, 99.51% F1-score, and a 0.42% false positive rate in the binary intrusion detection task. The study also shows that fea-ture selection is much more than just dimensionality-reduction; it is a very useful tool for enhancing the quality of alerts, an aspect that is often overlooked, in security operations. |
EngiTek 2026 |
Claim |
| 41 |
79.Cnf-1303 |
Anti-Money Laundering Analytics Using Blockchain Transaction Graph Intelligence |
Blockchain anti-money laundering analytics needs models that are capable of recognizing illicit transaction patterns without turning financial crime detection into a simple problem of classifying isolated tables. Therefore, detecting suspicious Bitcoin transactions is the main task of this proposed frame-work that leverages transaction-level features, temporal pay-ment-flow edges, neighborhood aggregation, and explainable risk scoring. The public Elliptic Bitcoin transaction graph is used as the experimental benchmark since it allows a realistic anti-money-laundering scenario with licit, illicit, and unknown transaction labels at different time points. The suggested method considers transactions as graph nodes and payment flows as the directed edges, then runs feature normalization, temporal splitting, class-imbalance handling, edge-aware rep-resentation learning, and investigator-oriented explanation. Experimental results reveal that graph-based learning im-proves the illicit-class F1-score and area under the precision-recall curve as compared to the non-graph baselines, especial-ly when suspicious behavior is spread over several hops and is not visible in a single transaction record. These results imply that blockchain AML programs should integrate machine learning, graph topology, security analytics, and decision-making support that is transparent rather than depending solely on traditional rule-based monitoring. |
EngiTek 2026 |
Claim |
| 42 |
79.Cnf-1281 |
Adversarially Robust AI Pipeline for Cybersecurity Analytics in Dynamic-Network Environments |
Dynamic networks generate large, diverse, and continuously changing cybersecurity telemetry. Hence, conventional attack-detection pipelines can be affected by distribution shifts, adversarial manipulation, and unstable decision boundaries. This article, therefore, proposes an adversarially robust artificial intelligence pipeline for cybersecurity analytics in dynamic networks. The pipeline combines telemetry normalization, feature-importance and stability-oriented inspection, adversarial training, adversarial robustness validation, drift-aware monitoring, explainable decision inspection, and operational feedback for security analysts. The study evaluates cyberattack classification under both clean and adversarially perturbed traffic conditions rather than focusing only on maximum conventional accuracy. The pipeline is evaluated using the UCI KDD Cup 1999 intrusion-detection dataset through standard classification metrics, adversarial degradation analysis, and robustness-retention comparison. The experimental results demonstrate that the adversarially trained gradient-boosting pipeline achieved strong clean-test performance while substantially reducing performance degradation under adversarial perturbation. The model recorded 98.14% clean accuracy, 96.88% clean F1-score, and 91.37% adversarial F1-score, thus exceeding the non-robust baseline under adversarial stress. The results suggest that robust cybersecurity analytics must integrate predictive performance, perturbation tolerance, and interpretability in order to provide trustworthy security operations in dynamic network environments. The experiment demonstrates the robustness behavior of the proposed pipeline on a benchmark intrusion-detection dataset, while full validation on modern cloud, IoT, encrypted, and enterprise SOC telemetry remains future work. |
EngiTek 2026 |
Claim |
| 43 |
79.Cnf-1287 |
Deep Learning-Based Detection of IoT Botnets in Smart Home Networks |
Smart home networks increasingly rely on Internet of Things (IoT) devices such as cameras, sensors, gateways, smart appli-ances, and voice-controlled assistants. However, weak authenti-cation, infrequent firmware updates, and limited local security controls create favorable conditions for botnet malware such as Mirai and BASHLITE. This article presents a deep learning-based detection framework for IoT botnets in smart-home net-works through an explainable artificial intelligence pipeline. The framework integrates traffic normalization, input-feature screen-ing, deep neural classification, threshold calibration, and SHAP/LIME-based interpretability to support accurate attack detection and understandable decision explanations. The pro-posed framework was evaluated using the N-BaIoT dataset, which contains network traffic collected from nine commercial IoT devices infected with Mirai and BASHLITE. The primary objective of the framework is to distinguish benign from mali-cious traffic while maintaining deployment feasibility on edge-assisted smart-home gateways. The study provides a structured detection pipeline, a clear evaluation procedure, and an inter-pretable reporting layer for improving smart-home cybersecurity resilience. |
EngiTek 2026 |
Claim |
| 44 |
79.Cnf-1284 |
Explainable Ensemble-Learning Decision Support for Resilient Cyber Defense |
Cyber defense teams are increasingly working in situations where the high number of alerts, uncertain evidence, different types of telemetry, and time constraints severely limit their capacity for making accurate decisions. This article proposes a new explainable ensemble learning-based decision-support framework to enable resilient cyber defense. This framework integrates a soft-voting ensemble classifier, a probability-based risk score, permutation-based global feature-importance analysis, and analyst-oriented decision tiers to facilitate the consistent prioritization of high-risk events. We believe that ensemble learning should be considered a division of transparency which could help security analysts by converting multidimensional evidences into explainable risk recommendations instead of replacing security analysts. The mechanism was then evaluated using domain-relevant data by building and testing a machine learning pipeline on the CICIDS2017 intrusion detection social dataset which contains labeled benign and malicious network-flow records, created in controlled traffic and attack scenarios. The held-out test partition results are 99.73%, 99.28%, 98.25%, 98.76%, and 99.94% on accuracy, precision, recall, F1 measure, and ROC-AUC respectively for the explainable soft voting ensemble. These results show that explainable ensemble learning combined with global feature-importance analysis, probability-based risk scoring, and human-in-the-loop escalation can contribute to the resilient cyber-defense decision processes. The experiment confirms the decision-support mechanism on a domain-aligned network-intrusion benchmark, but additional testing using security telemetry data from a live network and unknown attack scenarios is needed. |
EngiTek 2026 |
Claim |
| 45 |
79.Cnf-1306 |
Predictive Cyber-Risk Intelligence Model for Cross-Sector Digital Infrastructure |
As cross-sector digital infrastructure increasingly depends on interconnected cloud services, Internet of Things systems, transport platforms, public-service portals, financial workflows, and healthcare databases, the resulting interdependence expands the attack surface and creates a need for predictive cyber-risk intelligence models. This article presents a predictive cyber risk intelligence framework based on machine learning for identifying high-risk infrastructure states by way of benchmark security data and comparative model-performance analysis. The reason phishing is considered to be the representative cyber-risk benchmark for the study is that phishing remains one of the main entry points in attacks against organizational systems, public infrastructure, and service platforms. Therefore, the public UCI Phishing Websites dataset is used here. This study conducted a comparative experiment using classifiers Logistic Regression, Support Vector Machine, Random Forest, Gradient Boosting, and XGBoost. XGBoost achieved the strongest predictive performance among the evaluated models and achieved 97.42% accuracy, 97.61% precision, 97.22% recall, and 97.41% F1-score. The research demonstrates that ensemble-based learning methods, when integrated with operational decision-support mechanisms, can provide effective early-warning intelligence for cyber-risk assessment. This paper offers a workable predictive model that can be used by security analysts, infrastructure managers, and policymakers in cyber resilience planning. |
EngiTek 2026 |
Claim |
| 46 |
79.Cnf-1286 |
Explainable Deep Learning for Phishing Website Detection with Social-Engineering-Oriented Risk Interpretation |
Phishing sites are still developing and finding ways to exploit both technical and human vulnerabilities. The paper proposes an interpretable deep learning algorithm to detect phishing websites with structured URL, domain, page and trust-related features, and provides interpretable evidence to security analysts. The proposed model includes deep neural classification, feature attribution, confidence analysis and risk-layer mapping. The system is designed for cyber-resilient environments where automated detection needs to be used to support decision traceability, prioritization, and enforcement of the organization's policy. Experimental results for the UCI Phishing Websites dataset are used to benchmark logistic regression, random forest, gradient boosting, and multilayer perceptron against the proposed explainable deep learning model. The proposed model achieved 97.82% accuracy, 97.69% precision, 98.04% recall, a 97.86% F1-score, and 99.12% AUC. The explanation layer reveals important phishing cues like old domains, anchor-link mismatches, request-URL behavior, SSL state, and irregular URL structures. Based on the experimental results, the following can be concluded: (1) The explainable deep learning models can be used for phishing defense, as they have a strong phishing classification ability, in addition to providing the analysts with explainable support for their decisions. The evaluation takes into account only indicators of phishing via websites, as this is the dataset used to measure social-engineering risk. |
EngiTek 2026 |
Claim |
| 47 |
79.Cnf-1289 |
LLM-Guided Cyber-Defense Playbook Generation for SOAR-Enabled Environments |
Security orchestration, automation, and response (SOAR) tools are increasingly important in cybersecurity, requiring adaptive playbooks that convert threat intelligence into operational actions without introducing unsafe automation or excessive analyst burden. This paper proposes a playbook-generation framework for LLM-guided cyber defense in SOAR-enabled environments, where large language models assist in alert interpretation, evidence enrichment, response planning, and step-level containment logic. The prototype architecture integrates alert normalization, retrieval-augmented security context, policy-aware prompt control, human approval gates, and SOAR execution templates. The study uses the UCI phishing benchmark to evaluate the detection layer and a simulation-based prototype to assess how generated playbooks can support triage, prioritization, and containment recommendations. The machine-learning detection layer achieved strong phishing-classification performance, while the LLM-guided playbook layer improved analyst-oriented response coverage by transforming model outputs into structured response sequences. The results indicate that LLMs provide the greatest value in SOAR when their use is constrained by verified threat context, organizational policy, validation rules, and approval-based execution, rather than being deployed as autonomous incident-response agents. The experiment validates the phishing-detection and simulated playbook-generation components; full production SOAR deployment remains future work requiring live SOC alerts, real organizational policies, analyst feedback, and production SOAR platforms. |
EngiTek 2026 |
Claim |
| 48 |
79.Cnf-1291 |
Ensemble Machine Learning Model for Phishing Website Detection with Fuzzy Risk Indicators |
Phishing websites continue to pose a serious cybersecurity threat because they exploit both technical vulnerabilities and user trust while rapidly evolving their lexical, host-based, and content-level characteristics. This paper proposes a soft-voting ensemble machine-learning system augmented with fuzzy risk indicators that transform heterogeneous URL and webpage features into interpretable security-risk levels. The framework that has been proposed here brings together Random Forest, Gradient Boosting, Extra Trees, and Logistic Regression by means of a soft-voting ensemble, whereas a fuzzy indicator layer interprets chosen features into low, medium, high, and critical phishing-risk categories. The experiments were performed using a publicly available dataset of phishing websites that is based on the UCI/Kaggle phishing detection corpus. The findings reveal that the ensemble attained 96.81% accuracy, 96.42% precision, 97.15% recall, 96.78% F1-score, and 98.73% ROC-AUC, exceeding the performance of the separate baseline models. Additionally, the fuzzy risk layer greatly enhances the practical interpretability as it explains the model outputs via URL abnormality, domain trust, hyperlink behavior, and page integrity indicators. This paper shows that the integration of ensemble learning and fuzzy risk reasoning can improve phishing-detection performance while generating interpretable explanations for analysts, browser-warning systems, and organizational security gateways. |
EngiTek 2026 |
Claim |
| 49 |
80.Cnf-1264 |
Blockchain-Secured Edge Intelligence for Resilient Internet of Vehicles |
Internet of Vehicles (IoV) is emerging as an influential Cyber-Physical Network (CPN) in which vehicles, roadside equipment and infrastructure, edge servers, cloud services and intelligent decision-making systems collaborate to offer real-time mobility, safety and traffic management. However, false data injection, identity spoofing, privacy leakage, model poisoning, and unreli-able edge inferences are vulnerabilities of distributed connectivi-ty, to name a few. This paper proposes a Resilient IoV (RIoV) ecosystem framework by utilizing edge intelligence and block-chain. The proposed architecture conceptually integrates vehicle-side sensing, edge intelligence, federated learning, and trust management (based on blockchain), with the prototype experi-ments suggesting edge classification and simulated trust-layer validation. Implementation and testing use the UCI Statlog Ve-hicle Silhouettes dataset as a proxy benchmark for perception-driven edge classification. Gradient Boosting achieved the high-est classification accuracy of 79.3% and macro-F1 of 0.790, while Random Forest achieved a macro-F1 of 0.779 with lower average inference latency. Blockchain-enabled trust recording improves the traceability and accountability of inference deci-sions. The proposed framework therefore separates rapid edge inference from decentralized trust recording while emphasizing privacy-aware learning and lightweight edge analytics. |
EngiTek 2026 |
Claim |
| 50 |
80.Cnf-1278 |
Machine Learning Model for Insider Threat Detection Using Behavioral Log Patterns |
Insider threat detection is still one of the persistent challenges in enterprise cybersecurity, as the perpetrators are, in most cases, authenticated users and their activities do not differ much from legitimate ones. This paper presents a machine learning model for insider threat detection that exploits behav-ioral log patterns derived from authentication, device, file, email, web, and organizational activity traces. First, the frame-work extracts temporal behavioral indicators at the user level from raw user activity data. These indicators include abnormal login timing, device usage irregularity, file movement intensity, external communication frequency, web access deviation, and cross-channel risk accumulation. Then, a supervised learning pipeline is implemented that consists of feature normalization, class balancing, temporal aggregation, and model comparison across logistic regression, support vector machine, random forest, gradient boosting, and XGBoost classifiers. The experi-mental evaluation is based on the CERT Insider Threat Test Dataset whereby extracted daily user-level features are used to characterize normal and malicious behavior. The reported re-sults show that tree-based ensemble models provide strong predictive performance, with XGBoost obtaining the highest reported F1-score and ROC-AUC among the evaluated classifi-ers. The paper demonstrates that integrating different types of behavioral logs can aid the timely identification of insider threats, while offering transparent alerts for security operation centers. |
EngiTek 2026 |
Claim |
| 51 |
80.Cnf-1290 |
Lightweight Encryption and Authentication Scheme for Embedded IoT Security |
Embedded TCP/IP Internet of Things devices operate under severe constraints in memory, processing capacity, energy budget, and intermittently available network connectivity, while increasingly supporting safety-critical and privacy-sensitive functions. This paper proposes a lightweight encryption and authentication framework for securing resource-constrained embedded IoT devices. The framework integrates lightweight symmetric encryption, keyed message authentication, nonce-based freshness verification, and adaptive trust-aware session renewal. It is designed for resource-constrained sensor nodes, LAN gateways, and fog-assisted IoT deployments where con-ventional cryptographic mechanisms may introduce excessive latency, computational overhead, and energy consumption. The scheme was evaluated through a prototype-oriented experiment using traffic scenarios derived from the N-BaIoT dataset, which contains traffic from real IoT devices affected by Mirai and BASHLITE botnets. The evaluation indicates low authentication overhead while improving malicious-session rejection and maintaining a high benign-traffic acceptance rate. Key-renewal intervals and lightweight anomaly-triggered reauthentication are used to reduce unnecessary cryptographic operations. The resulting framework provides confidentiality, integrity, replay resistance, and deployable authentication without excessive resource consumption. |
EngiTek 2026 |
Claim |
| 52 |
80.Cnf-1300 |
Autonomous Threat Intelligence Correlation Using Knowledge Graphs and Explainable AI |
Automatically correlating threat intelligence evidence is one of the major challenges faced by modern security operations. This challenge arises because cyber incidents increasingly involve diverse evidence types, evolving attack chains, and scattered contextual signals across logs, alerts, vulnerabilities, malware reports, and external intelligence feeds. This article introduces an autonomous threat intelligence correlation framework that uses knowledge graph representation, machine-learning classification, and explainable artificial intelligence to support transparent cyber-threat reasoning. The proposed system transforms threat indicators into graph entities, connects them through semantic relationships, enriches them with contextual security features, and applies explainable learning models to rank malicious patterns. The framework aims to reduce analyst workload by connecting potentially overlooked observations into understandable threat narratives while maintaining traceability from model predictions to supporting evidence nodes. We evaluated the predictive classification layer using a public UCI/Kaggle phishing website benchmark containing 11,055 samples and 30 features. The obtained results reveal that the tree-based models achieved strong performance, with Extra Trees attaining 97.07% accuracy and 0.9741 F1-score. The results indicate that knowledge graph reasoning and explainable AI have the potential to empower cyber resilience through enhancing correlation accuracy, evidence visibility, and analyst trust. The experimental validation evaluates the predictive classification layer using a phishing benchmark dataset, while full validation of the knowledge graph and explainable correlation layers requires future testing on real multi-source threat intelligence telemetry. |
EngiTek 2026 |
Claim |
| 53 |
79.Cnf-1292 |
Secure Fog Computing Architecture for Real-Time Threat Detection in Smart Environments |
Smart environments are becoming heavily reliant on distributed sensors, gateways, edge nodes, and cloud services to support healthcare monitoring, smart buildings, industrial automation, energy control, and intelligent transportation. This growth not only enhances the service responsiveness but also raises the risk of attack as the entry point for malicious traffic can be a con-strained device, a compromised gateway, an insecure communi-cation link, and even vulnerable cloud interfaces. This article introduces a secure fog computing architecture designed to sup-port real-time threat detection in smart environments. The framework integrates local fog-layer preprocessing, a lightweight machine learning model, blockchain-assisted integrity logging, and policy-driven alert prioritization. Different from cloud-only intrusion detection, our architecture enables early detection near the data source, thus reducing latency, enhancing privacy, and limiting the propagation of attacks. The evaluation mainly re-volves around the public KDD Cup 1999 intrusion-detection dataset. It implements a reproducible machine-learning work-flow consisting of preprocessing, feature normalization, label encoding, class-weight adjustment, and comparative classifiers. On the KDD Cup 1999 benchmark, the proposed ensemble achieved 99.21% accuracy, 99.15% F1-score, and an average decision latency of 13.9 ms, compared with 48.6 ms for the eval-uated cloud-only configuration. This article provides a layered framework for resilient smart environments where cybersecurity decisions are distributed across sensing, fog, blockchain, and cloud intelligence layers. |
EngiTek 2026 |
Claim |
| 54 |
80.Cnf-1293 |
Explainable Large-Language-Model-Assisted Cyber Threat Intelligence for Resilient Security Operations |
Security Operations Centers currently struggle with a substantial increase in cyber alerts, diverse intelligence inputs, and limited analyst time for interpreting attack evidence. This paper pro-poses an interpretable large-language-model-assisted cyber threat intelligence framework for resilient security operations. The framework combines structured alert ingestion, feature-based threat scoring, natural language incident summarization, explainable reasoning, and analyst feedback into a single deci-sion-support pipeline. Instead of considering a large language model as an independent security decision-maker, the proposed architecture uses it as a transparent reasoning layer transform-ing diverse evidence into brief explanations, triage justifica-tions, and response recommendations. The NSL-KDD intrusion-detection benchmark was used to evaluate the structured classi-fication layer under controlled conditions. Its labeled network-connection records were mapped to Normal, DoS, Probe, R2L, and U2R classes. Logistic Regression achieved 79.72% accura-cy and a 68.00% macro F1-score, while the weaker R2L and U2R recall values confirmed the need for uncertainty-aware analyst review. |
EngiTek 2026 |
Claim |
| 55 |
80.Cnf-1294 |
Decentralized Identity Model for Privacy-Preserving Access Control in Web3 Systems |
Decentralized identity is becoming an important foundation for privacy-respecting access control in the Web3 space since cen-tral identity providers cannot fully provide complete user sover-eignty, selective disclosure, cross-platform interoperability, and auditable authorization. This paper presents an identity concept that integrates verifiable credentials, authentication via wallet, release of zero-knowledge attributes, policy enforcement with smart contracts, and federated risk-based adaptive access deci-sions. It is designed especially for Web3 users who wish to up-load only essential data to the application, intermediaries, or blockchain observers to establish their eligibility without divulg-ing unnecessary information. The proposed method separates the 4 main components of issuing identity, storing credentials, generating proofs, verifying policy, and risk estimation into distinct layers and avoids the collection of sensitive attributes, while ensuring accountability. The Kaggle Ethereum Transac-tion Fraud Detection dataset was used for experimental evalua-tion, which includes 85,003 transaction records and 14 features describing the behavior of transactions. The random forest verifier achieved 98.31% accuracy, a 98.31% F1-score, and a 0.9974 ROC-AUC, while the federated verifier reached 96.87% accuracy and reduced the estimated communication volume by approximately 99.91% relative to centralized raw-feature transfer. |
EngiTek 2026 |
Claim |
| 56 |
79.Cnf-1296 |
Data-Driven Cyber-Resilience Scoring Framework Using Explainable Machine Learning |
Cyber resilience requires more than isolated attack detection; it requires a quantifiable understanding of how organizational systems resist disruption, maintain essential functions, recover, and adapt after cyber incidents. This article proposes a data-driven cyber-resilience scoring framework that uses explainable machine learning to convert heterogeneous security indicators into a human-interpretable resilience index. Conceptually, the framework is designed to integrate vulnerability exposure, intrusion likelihood, authentication anomalies, service continuity, recovery readiness, and operational dependencies; however, the present experiment validates only the network-intrusion component using UNSW-NB15. A supervised machine-learning model is trained on cyber-risk-labeled observations, while interpretability methods identify the factors chiefly responsible for resilience degradation. The proposed approach is evaluated using the publicly available UNSW-NB15 cybersecurity dataset from Kaggle, which contains both normal and attack traffic categories suitable for resilience-oriented classification. Experimental results obtained using tree-based classifiers interpreted through SHAP-based feature-contribution analysis show that the proposed approach provides strong discrimination while retaining decision-level interpretability. The experiment validates the intrusion-risk scoring component of the proposed framework, while full cyber-resilience validation requires future testing with multi-source enterprise data covering vulnerability exposure, recovery readiness, service continuity, identity logs, and incident-response records. |
EngiTek 2026 |
Claim |
| 57 |
80.Cnf-1282 |
Imbalanced Intrusion Dataset Handling Using Sampling and Ensemble Learning |
Intrusion detection datasets are often overwhelmingly filled with normal traffic or very large classes of attacks. In contrast, the high-risk minority attacks occur rarely and are thus overlooked by the regular machine learning models. This article describes a sampling-aware ensemble learning approach to highly imbal-anced intrusion detection, demonstrated through a reproducible experiment on the UCI KDD Cup 1999 intrusion detection da-taset. The architecture of the framework includes preprocessing, categorical encoding, stratified partitioning, class imbalance cor-rection, and ensemble decision fusion that in turn results in better recognition of minority classes without compromising the overall detection reliability. Classifiers such as Random Forest, Gradient Boosting, and Balanced Bagging were tested against three differ-ent data-handling scenarios which were baseline training, ran-dom undersampling, and SMOTE-based oversampling. The re-sults demonstrate how imbalance-aware ensemble learning can take the macro-F1 value from 0.842 under the baseline Random Forest setting up to 0.917 with the proposed SMOTE-assisted ensemble configuration. Also, minority-class recall went up con-siderably, especially for Remote-to-Local and User-to-Root, which are both attack categories that occur rarely. The study reveals that mere accuracy is not a good enough measure to evaluate intru-sion detection systems when one is faced with a situation where class imbalance is very high. A thorough check with macro-F1, weighted-F1, minority recall, and false alarm rate yields a much better ground for picking the intrusion detection models that are ready for deployment. |
EngiTek 2026 |
Claim |
| 58 |
80.Cnf-1305 |
Federated Intrusion Detection for Distributed IoT Environments |
The operation of Distributed Internet of Things (IoT) environ-ments is progressively being extended through the use of heter-ogeneous sensors, gateways, edge nodes, fog clusters, and cloud services. This increases the attack surface that cannot be sufficiently covered by a single intrusion detection system (IDS). In this article, we intro-duce a federated intrusion detection system framework for dis-tributed IoT settings where, on the one hand, local edge clients independently train detection mod-els by using their local traffic data, and on the other hand, through a coordinating aggrega-tion layer, a global model is generated without the need to gath-er raw packet-level records. We believe that our proposed concept combines fed-erated learning, privacy-preserving update exchange, anomaly scoring at the edge level and model update audit through block-chain to provide enhanced detection resilience in distributed IoT envi-ronments. The Bot-IoT dataset on the Kaggle is used as the experimental bedrock. The Bot-IoT dataset available on Kaggle is employed as the experimental foundation. The outcomes pre-sented in the article compare centralized learning, local-only learning, and federated learning approaches under non-independent partitions of IoT traffic. The federated model achieved 97.63% accuracy and a 97.59% F1-score while elimi-nating the need to share raw traffic data and supporting scala-ble operation across distributed IoT environments. Among other things, this article offers a detailed system architecture, dataset-driven evaluation tables, and a discussion of how federated intrusion detection can be a support to cyber-resilient smart environments |
EngiTek 2026 |
Claim |
| 59 |
80.Cnf-1297 |
Deep Transfer Learning Model for Ransomware Behavior Detection in Endpoint Logs |
Ransomware behavior detection aims to identify behavioral indicators of a malicious attack on an endpoint early enough to reduce irreversible operational damage caused by encryption, persistence, lateral movement, or command execution. This article presents a deep transfer-learning model for ransomware-oriented behavior detection in endpoint logs using Windows API-call behavioral feature patterns as the primary evidence source and transferring learned malware-behavior representations to a target risk-classification task. Endpoint-log normalization, semantic behavior grouping, transfer-based deep feature extraction, attention-weighted feature modeling, and validation-based decision-threshold selection are combined in the proposed framework. The public UCI malware-behavior dataset used as the experimental basis provides Windows API-call observations for malware-family classification, with 7,107 instances and 280 features. The reported transfer model attained 98.21% accuracy, 98.06% precision, 98.44% recall, 98.25% F1-score, and 0.986 AUC, outperforming the evaluated random forest, support vector machine, multilayer perceptron, and non-transfer LSTM baselines. The results indicate that transfer learning may offer a promising basis for ransomware-oriented behavioral detection; however, robustness under scarce, noisy, and cross-environment endpoint data requires further validation. The paper also provides an interpretable and reproducible framework for ransomware-oriented behavior detection that can be integrated into endpoint detection and response pipelines, security operations centers, and cyber-resilience monitoring. |
EngiTek 2026 |
Claim |
| 60 |
79.Cnf-1299 |
Privacy-Preserving Cloud Data Sharing Framework for IoT-Enabled Enterprises |
IoT-enabled enterprises rely more and more on cloud platforms for their storage needs, analytics, collaboration, and gaining operational intelligence. However, sharing data directly on the cloud makes these enterprises vulnerable to privacy leakage, unauthorized inference, insider misuse, weak access governance, and the spread of attacks across edge-cloud channels. This arti-cle introduces a framework for cloud data sharing in a privacy-preserving manner for IoT-enabled enterprises. The proposed layered architecture integrates edge-level data minimization, attribute-based access control, encryption-aware sharing, threat scoring, and cloud-side auditability. The framework suits enter-prise environments where various sensors, gateways, applica-tions, and cloud services exchange sensitive operation data con-tinuously. The risk modeling portion of the proposed method, which identifies different streams of IoT data based on their confidentiality, sharing sensitivity, device trust, access purpose, and the likelihood of an attack, is the first in a series of steps. Before data is sent to cloud repositories, adaptive protection de-cisions are implemented. The security analytics part is assessed using the RT-IoT2022 dataset for attack categorization and risk scoring. The experimental results show that gradient boosting gave 98.71% accuracy and 98.42% macro-F1, whereas the pri-vacy-aware sharing layer lowered high-risk cloud exposure by 41.8% in the simulated enterprise workflow. The findings indi-cate that integrating privacy controls with IoT threat analytics can strengthen secure data sharing while preserving legitimate cloud-based collaboration. |
EngiTek 2026 |
Claim |
| 61 |
79.Cnf-1301 |
Blockchain-Based Cyber Resilience Framework for Secure Financial Transaction Monitoring |
Financial transaction monitoring systems increasingly rely on high-volume digital channels, real-time analytics, automated compliance checks, and data-driven fraud scoring. Even so, centralized monitoring platforms continue to be exposed to problems such as tampered audit logs, delayed anomaly con-firmation, fragmented institutional records, and weak ex-plainability during the post-incident investigation. This paper introduces a cyber resilience framework based on blockchain for secure financial transaction monitoring. The framework integrates permissioned blockchain logging, off-chain ma-chine-learning scoring, smart contract-based alert registra-tion, regulator-readable audit trails, and resilience-oriented recovery checkpoints. The proposed design is assessed using a transaction-risk classification process derived from the UCI Default of Credit Card Clients dataset, where payment behav-ior, credit utilization, delay patterns, and transaction history variables are used to detect suspicious or high-risk financial activities. The dataset is used as a public proxy for financial behavior risk classification rather than as a live banking fraud, blockchain, or institutional transaction-monitoring dataset. Experiments and simulations show that such an ap-proach to monitoring a permissioned blockchain is feasible by combining off-chain risk classification with tamper-evident alert registration, hash-based evidence verification and re-construction of the blockchain recovery state. The paper pro-vides a multi-layered structure, which integrates blockchain integrity, financial technology governance, and intelligent risk detection for the creation of secure digital finance envi-ronments. |
EngiTek 2026 |
Claim |
| 62 |
80.Cnf-1302 |
Risk Based Security Benchmarking for Cloud Edge IoT Deployment Models |
IoT deployment models such as cloud, edge, fog are increas-ingly integrated as unified cyber-physical infrastructuresBut it is difficult to make comparisons of security as each model apportions the tasks of computation, data, identity and moni-toring to different entities. This study proposes a risk-based security benchmarking framework for cloud-edge deployment models in IoT, which is based on a federated-learning evalua-tion design, the analysis of privacy-aware communication, and a multi-factor risk scoring mechanism. The research shows three deployment of IoT patterns: centralized cloud IoT, edge-assisted IoT, and fog-based federated cloud edge IoT. The proposed benchmark allows to quantify the detection perfor-mance, latency, communication cost, privacy exposure, and operational risk in the presence of adversarial network traffic. The experimental results are generated using a public IoT security dataset, and the obtained findings indicate that feder-ated cloud edge deployment strategy is the most balanced solution between attack detection and privacy preservation, although it entails a higher coordination overhead than cen-tralized cloud processing. The paper offers a workable benchmarking framework that can assist security architects, researchers, and deployment planners to decide among differ-ent IoT deployment options based on various factors rather than solely on accuracy |
EngiTek 2026 |
Claim |
|
|
