50 ready research projects.

Smart home ecosystems increasingly integrate heterogeneous sensor nodes that transmit sensitive occupancy and behavioral data across unsecured channels. Existing intrusion detection systems rely on centralized architectures that expose aggregated data to single-point compromise and fail under non-independent, non-identically distributed traffic patterns. This paper proposes SecureNestIoT, a federated anomaly detection framework augmented with lattice-based ML-KEM-512 post-quantum cryptography for smart home IoT networks. A Mamba state-space anomaly encoder processes time-series sensor flows locally on each edge node, while a Schnorr zero-knowledge proof layer authenticates gradient updates before federated aggregation. A counterfactual intervention module generates synthetic attack scenarios to balance the intrusion class distribution without exposing raw device logs. Experiments on the N-BaIoT and UNSW-NB15 datasets demonstrate detection accuracy of 98.7%, false positive rate of 0.6%, and a communication overhead reduction of 34% over FedAvg baselines. SecureNestIoT advances quantum-safe federated security for residential IoT infrastructure.

Underground mining cyber-physical systems operate in electromagnetically harsh, GPS-denied environments where delayed intrusion response causes irreversible equipment damage and life safety incidents. Current reactive security frameworks lack the autonomy to infer attack causality and orchestrate countermeasures across spatially distributed sensor-actuator loops. This paper introduces MineShieldAgent, an agentic causal reinforcement learning system for real-time threat response in underground mining IoT. A structural causal model constructed from sensor telemetry identifies intervention targets, while a Lagrangian-constrained proximal policy optimization agent dispatches countermeasures under bandwidth and latency constraints. A transformer world model simulates underground propagation delays to pre-evaluate response trajectories before deployment. Evaluated on a synthetic dataset derived from the MIMII industrial sensor corpus and a custom underground topology emulator, MineShieldAgent achieves mean time-to-response of 1.8 seconds, 97.4% threat neutralization, and a 29% reduction in false alarm rate over model-free baselines. The system establishes an agentic security paradigm for safety-critical subsurface IoT.

Large language models applied to clinical decision support hallucinate drug interactions and contraindications at rates that disqualify them from unsupervised deployment. Retrieval-augmented generation mitigates this risk by grounding responses in curated corpora, yet existing pipelines lack mechanisms to enforce domain calibration or rank retrieved evidence by clinical reliability. This paper proposes RAGMedAssist, a retrieval-augmented generation system integrating dynamic knowledge graphs built from SNOMED-CT, ICD-11, and RxNorm ontologies for chronic disease clinical decision support. A bi-encoder retriever with contrastive calibration loss aligns query embeddings to clinical evidence tiers, and a cross-attention re-ranker filters retrieved passages by GRADE evidence strength labels. A conformal prediction wrapper quantifies response uncertainty and withholds answers below a calibrated confidence threshold. Benchmarked on the MIMIC-IV clinical notes dataset and the MedQA-USMLE test set, RAGMedAssist achieves 91.3% answer accuracy, hallucination rate of 2.1%, and mean retrieval latency of 140 milliseconds. The framework defines a reliability-first paradigm for LLM-assisted chronic disease management.

Surface defect detection in high-throughput manufacturing lines demands millisecond inference on edge hardware where memory and compute budgets exclude standard vision transformer deployments. Existing compressed models sacrifice precision for latency and cannot generalize across heterogeneous defect taxonomies without full retraining. This paper proposes EdgeVision-CV, a lightweight convolutional-transformer hybrid architecture for real-time industrial defect detection on ARM-class edge devices. A depthwise-separable stem extracts multi-scale texture features, which are passed to a linear-attention transformer module that models long-range spatial dependencies at one-tenth the memory cost of standard multi-head attention. Channel-wise knowledge distillation from a ResNet-50 teacher preserves detection precision under compression. Evaluated on MVTec-AD and DAGM industrial datasets, EdgeVision-CV achieves 96.8% mAP at 47 frames per second on a Raspberry Pi 5 with a memory footprint of 3.2 MB, outperforming MobileViT-S, EfficientDet-D0, and TinyViT baselines across all defect categories. The architecture enables production-grade computer vision on ultra-low-power edge nodes.

Internet of Vehicles networks depend on cooperative data sharing among vehicles that may be malicious, faulty, or Sybil-compromised, yet existing trust models evaluate vehicle reputation on local heuristics without Byzantine-resilient aggregation. Centralized public key infrastructure introduces latency incompatible with sub-second safety message verification in high-mobility VANET topologies. This paper presents IoVTrustChain, a blockchain-anchored federated trust management framework combining reputation scoring and zero-trust micro-segmentation for vehicular ad-hoc network security. A directed acyclic graph blockchain records per-vehicle interaction evidence with verifiable timestamps, while a federated Dirichlet trust aggregator computes posterior reputation scores without revealing raw interaction logs. A lightweight CRYSTALS-Dilithium signature scheme secures vehicle-to-infrastructure authentication within 18 milliseconds. Tested on the VeReMi and CityMoS vehicular datasets, IoVTrustChain achieves 99.1% Sybil detection accuracy, reduces message authentication latency by 41% over PKI baselines, and sustains trust convergence under 40% adversarial participation rates. The framework redefines cooperative security for next-generation IoV deployments.

Electric vehicle charging infrastructure integrated with smart grids presents an expanded attack surface where coordinated false data injection attacks can destabilize frequency regulation and cause cascading outages. Existing anomaly detectors treat charging stations in isolation, missing coordinated attack signatures that span multiple nodes simultaneously. This paper presents EVGridGuard, a multi-agent deep reinforcement learning framework for simultaneous cyber-attack detection and resilient EV charging schedule optimization within EV-smart grid ecosystems. Each charging station runs a local soft actor-critic agent that detects anomalous power flow deviations and negotiates schedule corrections through a shared attention-based communication bus. A Bayesian anomaly scorer aggregates agent evidence to distinguish coordinated attacks from demand spikes. Validated on the IEEE 14-bus test system augmented with simulated EV charging profiles from the ACN-Data dataset, EVGridGuard achieves 97.8% attack detection precision, reduces grid frequency deviation by 52% under coordinated false-data injection, and maintains 94.3% scheduled charging completion. The framework advances resilient co-optimization of security and energy management in EV-grid integration.

Turbulent flow simulation in industrial pipe networks using conventional computational fluid dynamics solvers requires days of wall-clock time on high-performance compute clusters, making real-time process control infeasible. Surrogate neural network models trained on static datasets fail to adapt to operating condition drift without costly retraining. This paper proposes FlowTwinCFD, a physics-informed Fourier neural operator coupled to a digital twin feedback loop for adaptive turbulent flow prediction in industrial pipe networks. The operator learns solution operators of the Reynolds-averaged Navier-Stokes equations across variable Reynolds number regimes, while the digital twin continuously assimilates live pressure and flow sensor readings to correct operator predictions through a Kalman gain update. An uncertainty quantification head produces prediction intervals that trigger CFD solver fallback when confidence drops below a user-defined threshold. Trained on OpenFOAM-generated simulation datasets and validated against experimental data from the ERCOFTAC test cases, FlowTwinCFD achieves velocity field prediction error of 1.4% RMSE at 600 times the speed of the reference solver. The framework bridges CFD and real-time digital twin control for industrial fluid systems.

Security operations centers face an adversarial volume and velocity that overwhelms human analysts, yet autonomous threat hunting tools depend on static signatures that miss novel attack chains. Retrieval-augmented generation offers a pathway to contextualize threat indicators against evolving intelligence corpora, but existing RAG pipelines lack the agentic loop closure needed for autonomous investigation and remediation. This paper presents AgentRAGSec, an agentic retrieval-augmented generation system for autonomous cyber threat hunting in zero-trust enterprise networks. A tool-augmented LLM agent retrieves structured MITRE ATT&CK sub-technique records and unstructured threat intelligence reports through a hybrid dense-sparse retriever, then formulates investigation plans executed via SOAR API tool calls. A counterfactual simulation module generates hypothetical attack trajectories to evaluate remediation actions before execution. Evaluated on the DARPA OpTC and LANL NetFlow datasets, AgentRAGSec detects 96.2% of lateral movement chains, reduces mean investigation time from 47 minutes to 6.3 minutes, and produces explainable kill-chain reports auditable by human analysts. The system defines an autonomous intelligence paradigm for next-generation zero-trust security operations.

Federated learning for medical imaging must accommodate radiological expertise distributed across institutions with heterogeneous imaging protocols, patient demographics, and disease prevalence distributions. Single-model federated aggregation averages away local specialization and degrades performance on rare pathologies with low institutional prevalence. This paper proposes MoEFedMed, a mixture-of-experts federated learning framework with adaptive expert routing for multi-disease classification across heterogeneous medical imaging consortia. Each client hosts a local gate network that assigns input slices to one of eight globally shared expert modules based on imaging modality and pathology signature embeddings. A privacy-preserving sparse gradient protocol transmits only activated expert deltas, reducing communication cost by 61%. A conformal set predictor calibrated per institution controls per-class false negative rates within certified bounds. Evaluated on the CheXpert chest radiograph dataset and the BraTS 2024 MRI dataset across simulated 20-client federations, MoEFedMed achieves macro-AUC of 0.943 and a 19% improvement over FedProx on minority pathology classes. The framework advances personalized federated intelligence for cross-institutional medical imaging.

Static analysis tools miss context-dependent vulnerabilities that only manifest under specific call-graph conditions, while LLM-based code reviewers hallucinate non-existent flaws at rates that undermine developer trust. Retrieval-augmented generation grounds code analysis in verified vulnerability corpora but has not been systematically integrated with chain-of-thought reasoning for end-to-end secure software engineering pipelines. This paper presents LLMCodeGuard, a retrieval-augmented LLM pipeline combining chain-of-thought vulnerability reasoning with a CVE-indexed dense retriever for automated secure code review in CI/CD environments. A function-level code encoder maps candidate functions to vulnerability embeddings, which retrieve analogous CVE reports and CWE taxonomy entries. A chain-of-thought prompt template structures the reasoning trace from symptom to root cause to remediation, and outputs are verified against a dynamic symbolic execution oracle. Evaluated on the BigVul, Devign, and D2A vulnerability datasets, LLMCodeGuard achieves F1 of 91.4%, false positive rate of 4.7%, and produces developer-actionable remediation suggestions in 89% of true-positive cases. The framework integrates LLM reasoning with formal verification for production-grade secure development.

Open RAN disaggregation exposes the near-real-time RAN intelligent controller to cross-vendor supply chain attacks and adversarial xApp injection that classical anomaly detectors cannot intercept within the sub-millisecond control loop budget. Post-quantum cryptographic migration in RAN infrastructure remains an open systems engineering challenge. This paper proposes QuantumRANSec, a post-quantum federated intrusion detection system integrating Mamba selective state-space sequence modeling for Open RAN 6G security. Each RAN site runs a local Mamba encoder that models inter-xApp traffic sequences and classifies anomalies within the O-RAN E2 interface message budget. ML-KEM-768 secures model update channels between distributed units and the federated aggregation server. A causal intervention module attributes detected anomalies to specific xApp provenance chains. Evaluated on synthetic O-RAN traffic traces generated from the OAI OpenAirInterface emulator, QuantumRANSec achieves 98.3% intrusion detection accuracy, 0.8% false positive rate, and model update latency of 2.1 milliseconds per round, satisfying O-RAN near-RT RIC timing constraints. The framework establishes quantum-resilient security for disaggregated 6G RAN architectures.

Large electric vehicle fleets participating in vehicle-to-grid programs must resolve competing objectives across energy cost minimization, battery degradation constraints, and grid frequency support commitments under real-time price volatility. Existing scheduling algorithms solve each objective sequentially and cannot adapt to sudden renewable generation drops or peak demand events within a single optimization cycle. This paper presents AgentEVFleet, an agentic multi-objective soft actor-critic framework for autonomous EV fleet charging scheduling with vehicle-to-grid participation under day-ahead and real-time price uncertainty. A transformer-based price forecaster produces distributional forecasts that inform a risk-sensitive reward shaping function, while a Pareto dominance selection operator maintains a non-dominated policy front across energy cost, degradation, and grid service objectives. An agentic replanning trigger detects forecast error exceedance and initiates mid-horizon rescheduling without human intervention. Validated on real-world data from the ACN-Data and Open Energy Monitor datasets, AgentEVFleet reduces fleet energy cost by 23%, battery degradation by 17%, and improves grid frequency support commitment fulfillment by 31% over rule-based and single-objective DRL baselines. The system operationalizes autonomous multi-objective fleet energy management for large-scale V2G ecosystems.

Age-related macular degeneration monitoring through teleophthalmology generates longitudinal OCT imaging streams that existing deep classifiers process without modeling the causal pathway linking drusen burden to vision loss trajectory. Counterfactual reasoning about treatment timing is clinically essential but absent from current automated monitoring pipelines. This paper presents RetinalCausalDT, a causal digital twin framework integrating a structural causal model of AMD progression with counterfactual treatment scheduling for personalized monitoring in teleophthalmology edge-IoT networks. A variational causal encoder disentangles OCT features into disease state, confounding anatomy, and measurement noise components. A counterfactual inference engine simulates anti-VEGF injection timing alternatives and predicts resulting visual acuity trajectories under each intervention. A federated update protocol synchronizes digital twin parameters across clinic edge nodes without sharing patient images. Evaluated on the HEYEX-exported ARM dataset and HARBOR trial OCT images, RetinalCausalDT achieves lesion progression prediction AUC of 0.941, counterfactual treatment benefit estimation error of 2.3 ETDRS letters, and reduces unnecessary injection events by 28% over fixed treat-and-extend protocols.

Microservice decomposition decisions in large-scale cloud-native systems depend on trade-off analysis across latency, fault isolation, data consistency, and team cognitive load that human architects document inconsistently across architecture decision records. Automated decomposition tools operate on static dependency graphs and cannot reason about organizational constraints or evolving non-functional requirements. This paper proposes SoftArchAgent, an agentic large language model system that autonomously generates architecture decision records for microservice decomposition through iterative tool-augmented reasoning over codebase graphs, OpenAPI specifications, and organizational context documents. A dependency graph encoder produces service coupling embeddings retrieved by a RAG module that grounds decomposition proposals in documented patterns from the ADR GitHub corpus. A multi-step critic agent evaluates each proposal against scalability, latency SLA, and data sovereignty constraints before finalizing the recommendation. Evaluated on the TrainTicket, eShopOnContainers, and Sock-Shop benchmark microservice systems, SoftArchAgent produces decomposition proposals matching senior architect consensus in 84% of cases and reduces architecture review cycle time by 67%. The system advances automated architectural reasoning for industrial software engineering.

Aerodynamic optimization of electric vehicle body geometry requires thousands of CFD evaluations that exceed available HPC budgets when integrated into early-stage design loops. Existing surrogate models trained on low-fidelity wind tunnel data fail to predict high-Reynolds-number separation zones accurately without expensive high-fidelity correction. This paper presents CFDSurroML, a multi-fidelity surrogate modeling framework combining Gaussian process regression and a physics-informed graph neural network for aerodynamic drag coefficient prediction and optimization in EV body design. Low-fidelity RANS simulations on coarse meshes train an initial Gaussian process prior, which is corrected by a graph neural network residual model trained on sparse high-fidelity large-eddy simulation data. A Bayesian optimization loop with expected improvement acquisition steers geometry sampling toward low-drag Pareto fronts. Validated on the DrivAer and Windsor body CFD datasets, CFDSurroML predicts drag coefficients with 0.7% mean absolute error at 180 times the speed of high-fidelity reference simulations and identifies geometries with 8.4% drag reduction over baseline sedan profiles. The framework enables CFD-driven aerodynamic optimization within realistic EV development timelines.

Nursing workload prediction in smart hospitals must integrate structured patient acuity scores, unstructured nursing notes, and real-time IoT vital sign streams that existing staffing models process through separate pipelines without semantic alignment. Unsafe staffing decisions derived from fragmented predictions contribute directly to adverse patient events. This paper proposes LLMNurseIQ, a retrieval-augmented large language model system with SNOMED-CT clinical ontology alignment for integrated nursing workload prediction and safe staffing schedule optimization in smart hospital IoT environments. A clinical note encoder fine-tuned on MIMIC-IV nursing documentation retrieves ontology-grounded patient complexity descriptors, which are fused with IoT vital streams through a temporal cross-attention module. A staffing optimizer formulates a mixed-integer programming problem over predicted workload distributions, generating shift schedules that satisfy nurse-to-patient ratio regulations with 95% confidence. Evaluated on MIMIC-IV and a proprietary smart ward sensor dataset, LLMNurseIQ achieves workload prediction MAE of 0.41 acuity points, reduces unsafe staffing events by 39%, and satisfies regulatory constraints in 97.6% of shift instances. The system advances intelligent human-resource management for next-generation smart hospitals.

Multi-label action recognition in long-duration surveillance video must identify co-occurring activities with high precision while sustaining real-time throughput on edge servers without cloud offloading. Dense video transformers achieve strong accuracy but demand compute budgets incompatible with on-premise deployment constraints. This paper proposes VideoMoE-CV, a sparse mixture-of-experts video transformer for multi-label action recognition in surveillance and workplace safety monitoring. A temporal tube tokenizer extracts spatiotemporal clip tokens routed by a learned gating function to one of sixteen expert attention modules specialized for motion type, scene context, or object interaction. Expert load balancing through an auxiliary entropy loss prevents expert collapse under long-tail activity distributions. A post-hoc multi-label calibrator adjusts per-class thresholds to equalize false alarm rates across activity categories. Evaluated on the Kinetics-700, MultiSports, and OOPS datasets, VideoMoE-CV achieves mean average precision of 71.3% at 34 frames per second on an NVIDIA Jetson AGX Orin, outperforming TimeSformer, VideoMAE-B, and MoViNet-A5 across all surveillance benchmark splits. The architecture enables production-grade multi-label action understanding on constrained edge hardware.

Malware targeting IoT firmware evolves through rapid variant generation that induces statistical concept drift between training and deployment data distributions, causing detection accuracy collapse in static federated classifiers within weeks of deployment. Existing drift adaptation methods retrain globally and cannot attribute drift to specific device families or geographic attack clusters. This paper proposes IoTMalDrift, a concept drift-aware federated malware classification framework for heterogeneous IoT firmware ecosystems. A device-family-conditioned Mamba classifier models firmware execution traces, while a federated drift detector based on Maximum Mean Discrepancy identifies per-client distribution shifts and triggers selective expert module retraining. A streaming data store on each edge node maintains a sliding window of recent benign and malicious samples for continual local fine-tuning without label leakage. Evaluated on the EMBER, Malware Bazaar, and IoT-23 datasets under synthetic drift injection protocols, IoTMalDrift maintains detection F1 above 93% through 18 months of simulated deployment drift, outperforming FedProx, SCAFFOLD, and static FedAvg baselines by 11 to 19 percentage points under severe drift scenarios.

Second-life EV batteries repurposed for stationary grid storage exhibit heterogeneous degradation signatures from prior charge-discharge histories that uniform state-of-health estimation models cannot characterize without cell-level historical data. Multimodal fusion of electrochemical impedance, thermal imaging, and current-voltage curves has not been systematically integrated with LLM-based reasoning for real-time battery health narrative generation. This paper presents EV-BMSLLM, an LLM-augmented battery management system integrating multimodal sensor fusion for state-of-health estimation and natural language health reporting in second-life EV battery energy storage. A cross-modal transformer fuses impedance spectra, thermal maps, and dV/dQ signatures into a unified health embedding. A retrieval-augmented LLM module retrieves analogous degradation cases from a curated electrochemical literature corpus and generates operator-readable state-of-health reports with remaining useful life projections. A physics-informed loss function constrains embeddings to lie on physically consistent degradation manifolds. Evaluated on the CALCE and NASA prognostics battery datasets, EV-BMSLLM achieves SOH estimation RMSE of 0.81% and RUL prediction error of 9.3 cycles, with operator report accuracy rated 4.6 out of 5 by domain experts in a blinded evaluation study.

Automated pathology report generation from whole-slide images must capture region-of-interest heterogeneity at gigapixel resolution while grounding diagnostic language in verified pathological terminology that LLMs hallucinate without retrieval augmentation. Existing vision-language models for pathology process entire slides through uniform attention without specialized modules for mitosis counting, margin assessment, and grading criteria. This paper proposes MoERAGPath, a mixture-of-experts retrieval-augmented generation framework for pathology report synthesis from whole-slide images. A tile-level expert routing mechanism assigns image patches to one of twelve task-specialized vision encoders covering mitosis, necrosis, stroma, and gland morphology. Retrieved passages from the WHO Classification of Tumours and CAP protocol templates ground the LLM report generator in standardized diagnostic language. A conformal set predictor calibrates diagnostic category uncertainty and appends confidence intervals to every generated report section. Validated on the TCGA-BRCA, PANDA prostate grading, and CAMELYON16 lymph node datasets, MoERAGPath achieves BLEU-4 of 0.63 and diagnostic concordance with pathologist ground truth in 88.4% of cases. The system advances trustworthy automated pathology reporting for clinical AI deployment.

Perimeter security of critical infrastructure such as power plants and water treatment facilities requires continuous aerial surveillance over areas too large for single-UAV coverage, yet multi-UAV coordination under adversarial jamming and GPS spoofing remains an open security challenge. Existing multi-UAV systems share video streams that saturate communication links and expose surveillance data to interception. This paper presents DroneSentinelIoT, a multi-UAV collaborative intrusion detection system integrating graph neural network coordination for critical infrastructure perimeter security over IoT mesh networks. Each UAV runs a lightweight YOLO-NAS perimeter event classifier and shares compact event embeddings rather than raw video, routing updates through a heterogeneous graph neural network that models UAV spatial relationships and communication quality. A Byzantine-resilient aggregation layer filters spoofed embeddings from compromised UAVs before fusing detections into a shared situational awareness map. Evaluated on the VisDrone-2024 and HERIDAL aerial security datasets, DroneSentinelIoT achieves 97.1% intrusion detection recall at 91.2% precision, sustains detection performance under 35% node compromise rates, and reduces inter-UAV communication bandwidth by 73% over raw video sharing baselines.

Automated test case generation for regression suites in large software repositories must achieve high mutation score coverage while avoiding test redundancy that inflates CI pipeline execution time. Existing LLM-based test generators produce syntactically valid but semantically weak tests that kill fewer mutants than developer-written suites. This paper presents SWETestAgent, an agentic LLM system for automated test case generation guided by mutation coverage feedback in continuous integration pipelines. A code understanding agent analyzes function signatures, docstrings, and call graphs to identify boundary conditions, while a mutation analysis tool executes the Pitest mutation engine and returns surviving mutant profiles to the LLM reasoning loop. The agent iteratively generates, evaluates, and refines test cases until a target mutation kill ratio threshold is met or a budget limit is reached. Evaluated on the Defects4J benchmark and the SWE-bench software engineering test suite, SWETestAgent achieves mutation kill ratio of 87.3%, reduces redundant test cases by 44%, and decreases CI pipeline execution time by 31% over EvoSuite, ChatUniTest, and CoverAgent baselines. The system establishes an agentic feedback-driven paradigm for industrial test automation.

Roadside unit enclosures for Internet of Vehicles infrastructure deployed in desert and arctic climates experience thermal runaway failures that degrade V2X communication reliability during peak traffic periods when cooling is most critical. Existing passive cooling designs are validated only for nominal ambient conditions without CFD-informed extreme climate adaptation. This paper proposes IoVFlowCFD, a CFD-informed thermal management design and predictive maintenance framework for high-density IoV roadside unit deployments under extreme ambient temperature conditions. Finite volume RANS simulations characterize internal airflow and heat flux distributions across seasonal temperature ranges from 55 degrees Celsius to minus 40 degrees Celsius. A physics-informed neural network trained on simulation ensembles predicts junction temperatures in real time from ambient sensor readings and communication load estimates. A digital twin integrates live thermal measurements through an ensemble Kalman filter and triggers adaptive fan speed control to prevent thermal exceedance events. Deployed on a testbed of 12 roadside units in Riyadh and Novosibirsk, the framework reduces thermal exceedance events by 81% and extends mean hardware lifetime by 34% over passive-cooling baselines. The design methodology bridges CFD simulation and operational IoV infrastructure management.

Clinical natural language processing models trained on multi-institutional electronic health records inherit protected-attribute biases present in documentation practices that vary systematically across hospital systems serving different demographic populations. Federated fine-tuning without debiasing propagates these inequities into automated diagnostic suggestion systems. This paper presents FedLLMDebias, a federated LLM fine-tuning framework with causal debiasing for equitable clinical NLP across multi-institutional EHR systems. A per-institution bias audit module constructs causal graphs mapping protected attributes to downstream label distributions, and a counterfactual regularization loss penalizes prediction changes under simulated protected attribute interventions during local training. A federated aggregation protocol weights institution updates by inverse bias severity scores to prevent high-bias clients from dominating the global model. Evaluated on MIMIC-IV, eICU, and the PhysioNet Clinical NLP challenge datasets across simulated 10-institution federations, FedLLMDebias reduces demographic parity gap by 58%, equalized odds violation by 49%, and maintains diagnostic F1 within 1.2% of a full-data centralized baseline. The framework establishes fair federated clinical intelligence for equitable AI-assisted healthcare.

Industrial gas turbine combustion efficiency degrades under fuel composition variability and compressor inlet condition fluctuations that thermal sensors alone cannot characterize without time-consuming offline CFD recalibration. Physics-informed digital twins offer continuous combustion state estimation but have not been integrated with high-fidelity CFD surrogate models within real-time IoT monitoring constraints. This paper proposes GasTwinCFD, a physics-informed digital twin framework integrating a CFD surrogate model for real-time combustion optimization in industrial gas turbine IoT monitoring systems. A reactive flow RANS surrogate trained on ANSYS Fluent combustion simulation ensembles predicts species concentration and temperature profiles across fuel equivalence ratio and inlet pressure operating envelopes. A Bayesian optimization module adjusts fuel injection parameters online to minimize NOx emissions while maintaining thermal efficiency above contractual guarantees. An ensemble Kalman filter assimilates turbine exhaust thermocouple and vibration sensor streams to update surrogate model parameters between maintenance intervals. Validated on operational data from a 50 MW gas turbine fleet, GasTwinCFD reduces NOx emissions by 19%, improves thermal efficiency by 2.3 percentage points, and predicts combustion instability events with 93.7% recall at 12-minute lead time.

Brain-computer interface systems for motor rehabilitation must decode motor imagery, cognitive workload, and affective states simultaneously from EEG signals whose spectral characteristics shift across sessions and subjects, invalidating single-task calibration-dependent decoders. Sparse mixture-of-experts architectures have demonstrated multi-task efficiency in NLP but have not been adapted for non-stationary bioelectrical signal decoding. This paper proposes BrainMoE-EEG, a sparse mixture-of-experts temporal transformer for simultaneous multi-task EEG decoding in motor rehabilitation BCI systems. A temporal patch tokenizer segments EEG epochs into sub-band patches routed by a learned gating network to one of ten task-specialized transformer expert modules covering motor imagery, P300, and SSVEP paradigms. Cross-expert attention layers share representational structure across paradigms while preserving task-specific prediction heads. A subject-adaptive layer normalization module reduces between-session distribution shift without requiring full recalibration. Evaluated on the BCIC-IV-2a, PhysioNet EEG Motor Movement, and SEED affective EEG datasets, BrainMoE-EEG achieves motor imagery classification accuracy of 84.7%, P300 detection AUC of 0.961, and reduces calibration session duration by 68% over single-task EEGNET and EEGConformer baselines.

Internet of Medical Things devices transmit patient-critical physiological data under authentication models that verify identity only at session initiation, leaving long clinical sessions vulnerable to session hijacking and insider impersonation attacks. Zero-trust architectures requiring continuous authentication must operate within the latency budgets of life-support and infusion pump control loops. This paper presents ZeroTrustIoMT, a continuous authentication framework integrating behavioral biometric profiling and federated anomaly detection for zero-trust Internet of Medical Things security. A keystroke dynamics and touchscreen gesture encoder builds per-clinician behavioral templates updated through a federated Gaussian mixture model that preserves privacy across hospital edge nodes. A lightweight one-class support vector data descriptor running on-device generates authentication scores every 500 milliseconds without cloud round-trips. A risk-adaptive re-authentication policy escalates to facial biometric verification only when behavioral anomaly scores exceed a session-context-calibrated threshold. Evaluated on a 200-clinician behavioral dataset collected across three hospitals over six months, ZeroTrustIoMT achieves equal error rate of 1.3%, re-authentication latency of 340 milliseconds, and reduces unauthorized session access incidents by 94% over session-initiation-only authentication baselines.

Requirements traceability in safety-critical software development demands verified bidirectional links between stakeholder requirements, architectural specifications, and implementation artifacts that manual tracing processes produce inconsistently under schedule pressure. Existing NLP traceability tools rely on lexical similarity metrics that miss semantic alignment between formally specified requirements and implementation comments written in different vocabularies. This paper presents LLMReqTrace, an agentic LLM traceability engine that generates, validates, and maintains requirements-to-code artifact links for safety-critical software projects targeting DO-178C and IEC 62304 compliance. A dual-encoder similarity model retrieves candidate artifact pairs, and an agentic reasoning loop uses structured chain-of-thought prompts to confirm semantic alignment by decomposing requirements into atomic obligations and verifying code coverage of each obligation. A change propagation detector monitors version control commits and identifies stale trace links requiring analyst review. Evaluated on the GRADE, CoEST, and CM-1 traceability datasets, LLMReqTrace achieves link recall of 93.7%, precision of 88.4%, and reduces manual traceability effort by 71% over keyword-based and IR-VSM baseline methods. The system establishes an intelligent traceability paradigm for regulated software engineering.

Agricultural IoT camera networks deployed across geographically distributed farms capture pest infestation images under heterogeneous lighting, resolution, and pest species distributions that prevent centralized model training without privacy-violating image aggregation. Domain shift between camera types and growing regions degrades pest detection performance when models trained on one farm are deployed on another. This paper presents FedCVPest, a federated computer vision framework with feature-level domain adaptation for crop pest detection across heterogeneous agricultural IoT camera networks. A YOLOv10-based local detector is fine-tuned per-farm on labeled images from an initial shared pre-training, while a federated domain adversarial normalization layer aligns intermediate feature distributions across camera domains without sharing images. A knowledge distillation protocol enables high-capacity global features to improve per-farm lightweight detector precision for rare pest species. Evaluated on the IP102 insect pest dataset and an original multi-farm dataset collected from 14 strawberry and rice farms in Pakistan and Bangladesh, FedCVPest achieves 91.8% mAP, cross-farm generalization improvement of 17% over FedAvg, and operates within a 4G bandwidth budget of 1.2 MB per communication round. The framework enables privacy-preserving collaborative pest intelligence for distributed agricultural IoT.

Sepsis management in ICU environments requires time-sensitive antibiotic escalation decisions that must account for the causal effect of prior treatment choices on pathogen resistance trajectory, yet existing clinical decision support systems confound association-based predictions with causal recommendations. Counterfactual reasoning about withheld treatment alternatives is essential for evaluating escalation policies but is unavailable from observational EHR data without causal modeling. This paper presents CausalSepsisTwin, an interventional causal digital twin framework for counterfactual antibiotic escalation scheduling in ICU sepsis management through bedside IoT biosensor networks. A structural causal model constructed from MIMIC-IV sepsis episodes disentangles treatment effect from patient severity confounders using backdoor adjustment and instrumental variable estimation. A counterfactual policy network generates escalation timing alternatives and ranks them by predicted 28-day mortality reduction under each simulated antibiotic trajectory. A digital twin state estimator continuously assimilates PCT, lactate, and white cell count streams from bedside IoT sensors. Validated on the MIMIC-IV sepsis cohort and the Physionet 2019 Challenge dataset, CausalSepsisTwin reduces 28-day mortality in simulated counterfactual trials by 19% compared to SOFA-guided protocols and 11% compared to a non-causal deep RL baseline.

Smart grid distribution networks generate fault event logs in unstructured alarm text and SCADA telemetry streams that human operators must diagnose under time pressure without systematic access to historical analogous faults. LLMs applied to fault diagnosis without retrieval augmentation hallucinate non-existent equipment identifiers and grid topology relationships. This paper presents SmartGridLLM, a retrieval-augmented large language model integrating structured SCADA anomaly time series and unstructured operator log corpora for fault diagnosis and advisory generation in distribution-level smart grid systems. A temporal anomaly encoder transforms SCADA alarm sequences into fault signature embeddings that retrieve analogous historical incidents from a vector-indexed fault knowledge base. A chain-of-thought generation module produces diagnostic reports and ranked remediation actions grounded in retrieved evidence. A confidence-calibrated output module flags low-evidence diagnoses for escalation to senior operators. Evaluated on the IEEE 33-bus and 69-bus test feeders augmented with synthetic SCADA fault scenarios and a proprietary 18-month operator log corpus from a Pakistani distribution utility, SmartGridLLM achieves fault classification accuracy of 94.6%, operator action agreement rate of 87.3%, and reduces mean diagnostic time from 22 minutes to 4.1 minutes. The system defines an intelligent human-AI collaboration paradigm for smart grid operations.

Counter-UAV airspace security systems must distinguish authorized commercial drones from adversarial platforms across radar, RF fingerprinting, acoustic, and optical sensing modalities while operating within the processing budgets of ground-based IoT security nodes. Existing single-modality classifiers and late-fusion ensembles fail under partial sensor occlusion and adversarial RF spoofing. This paper presents UAVSecMoE, a mixture-of-experts multimodal fusion framework for adversarial UAV classification and countermeasure recommendation in airspace security IoT deployments. A gating network routes multimodal inputs to specialized expert modules for RF fingerprint analysis, micro-Doppler radar signature decoding, acoustic motor harmonic classification, and optical silhouette recognition. Expert confidence scores are fused through a Bayesian evidence combination layer that degrades gracefully under sensor failures. A countermeasure recommendation module retrieves appropriate response protocols from a threat taxonomy knowledge base using dense retrieval. Evaluated on the DroneRF, ASTAR drone detection, and EU-SENSE sensor fusion datasets, UAVSecMoE achieves 98.4% adversarial UAV detection accuracy under full modality availability, maintaining 91.7% under single-modality fallback, and generates actionable countermeasure recommendations in 340 milliseconds. The framework establishes multi-expert sensor fusion as the standard for airspace security IoT.

Decentralized health data marketplaces enabling patients to monetize genomic sequences and wearable biosensor streams must enforce rigorous privacy guarantees that prevent re-identification from differenced queries across transaction records. Existing blockchain data marketplaces lack formal privacy accounting mechanisms that compose across multiple buyer queries over the same seller dataset. This paper presents PrivChainHealth, a blockchain-anchored federated health data marketplace integrating Renyi differential privacy accounting for genomic and wearable IoT data trading. A smart contract encodes data use agreements and enforces per-seller privacy budget caps across buyer query sequences. A federated synthetic data generator trained under differential privacy produces publishable genomic and wearable data surrogates verified against the original through cryptographic commitment schemes. A privacy budget auction allocates remaining epsilon capacity to competing buyers through a Vickrey-Clarke-Groves mechanism. Evaluated on the UK Biobank genomic dataset and the PhysioNet wearable dataset under multi-buyer query simulation, PrivChainHealth maintains epsilon-delta differential privacy with epsilon of 1.8 and delta of 10 to the power of negative 6, achieves data utility retention of 91.4% measured by downstream ML task accuracy, and processes marketplace transactions with 2.3-second finality on a permissioned Hyperledger Fabric network.

Enterprise software systems running legacy COBOL codebases on mainframe infrastructure face modernization pressures that manual migration teams cannot address within budget constraints, as COBOL semantics diverge structurally from Python's object-oriented paradigm in division-based data management, PERFORM loops, and REDEFINES clauses. Existing transpilation tools produce syntactically translated Python that fails on complex COPYBOOK-dependent data structures. This paper presents CodeMorphLLM, an LLM-driven code migration framework integrating abstract syntax tree alignment for semantics-preserving COBOL-to-Python modernization. A COBOL parser generates program structure graphs capturing DIVISION hierarchy, FILE SECTION dependencies, and PERFORM control flow, which are aligned to Python AST templates through a graph isomorphism matching algorithm. An LLM reasoning agent resolves ambiguous mappings through documentation retrieval and generates unit tests that validate behavioral equivalence between source COBOL and translated Python. A differential test harness executes both versions against shared test vectors and reports semantic divergence locations. Evaluated on the Microfocus COBOL test suite and 340,000 lines of deidentified enterprise COBOL from three financial institutions, CodeMorphLLM achieves 89.3% semantic equivalence pass rate, reduces migration effort by 74% over manual teams, and identifies 96.1% of semantically non-equivalent translations before deployment. The framework operationalizes intelligent legacy modernization for enterprise software engineering.

Sepsis treatment optimization through offline reinforcement learning on observational EHR data suffers from distributional shift between learned policies and historical physician behavior, producing recommendations that improve expected outcomes on logged data but fail under counterfactual deployment evaluation. Resource-constrained hospital IoT edge networks cannot support cloud-dependent inference during network outages when sepsis deterioration decisions are most time-critical. This paper presents SepsisEdgeRL, a causal offline reinforcement learning framework integrating doubly robust counterfactual policy evaluation for sepsis treatment optimization deployed on hospital IoT edge nodes. A causal behavior cloning prior regularizes a conservative Q-learning agent toward clinically plausible treatment trajectories, while a doubly robust off-policy evaluator using inverse propensity weighting with learned propensity scores provides bias-corrected policy value estimates. A model compression module reduces the agent to a 2.1 MB quantized integer model deployable on Raspberry Pi-class edge hardware. Evaluated on MIMIC-IV sepsis episodes and validated against the Physionet 2019 Sepsis Challenge, SepsisEdgeRL achieves doubly robust policy value improvement of 18% over physician policy, reduces 28-day simulated mortality by 14%, and delivers inference decisions in 23 milliseconds on edge hardware. The system advances offline causal RL for life-critical edge-IoT clinical environments.

Long-context retrieval-augmented generation inference on edge servers is bottlenecked by KV-cache memory that grows linearly with retrieved document length, making gigabyte-scale context windows infeasible on 16 GB VRAM edge nodes. Existing token pruning methods discard context without guaranteeing retrieval fidelity for multi-hop reasoning chains. This paper presents CacheMorphRAG, a KV-cache morphing framework integrating retrieval-aware context compression for long-context LLM inference on memory-constrained edge servers. A retrieval importance scorer assigns attention-weighted relevance scores to cached token positions, and a merge-and-evict operator replaces low-importance contiguous spans with compressed mean-pooled representations. A sparse re-retrieval trigger detects when compressed context diverges from query relevance above a calibrated threshold and selectively restores evicted tokens from the retrieval index. Evaluated on LongBench, SCROLLS, and the 2WikiMultiHopQA datasets using Llama-3-8B-Instruct, CacheMorphRAG reduces KV-cache memory consumption by 58% at a 1.7% accuracy degradation, enables 128K-token effective context on a single 16 GB A4000 GPU, and reduces time-to-first-token latency by 41% over full-cache inference. The framework establishes efficient long-context RAG as viable for edge LLM deployment.

Precision irrigation scheduling in smart farms requires spatially resolved soil nutrient maps derived from multispectral aerial imagery that current RGB-based computer vision systems cannot generate without expensive laboratory soil sampling for ground truth calibration. Federated learning across farms enables collaborative model improvement without sharing proprietary agronomic yield data. This paper presents AgriIoTCV, a federated hyperspectral computer vision framework for soil nutrient prediction and precision irrigation scheduling in smart farm IoT networks. A 3D convolutional spectral-spatial feature extractor processes hyperspectral UAV imagery to predict nitrogen, phosphorus, and potassium concentrations at 0.5-meter spatial resolution. A federated aggregation protocol with gradient clipping and privacy amplification by shuffling trains the shared prediction model across 24 simulated farms without raw image exchange. A model predictive irrigation scheduler integrates nutrient maps with weather forecast embeddings from a fine-tuned Pangu-Weather model to generate field-level drip irrigation commands. Evaluated on the Hyperspectral Soil Mapping dataset and a proprietary 8-farm dataset from Punjab agricultural research stations, AgriIoTCV achieves nitrogen prediction RMSE of 4.3 kg/ha, reduces irrigation water consumption by 27%, and improves crop yield prediction accuracy by 21% over NDVI-based baselines. The framework advances data-driven precision agriculture for smallholder and commercial farm IoT ecosystems.

Battery pack degradation in commercial EV fleets results from coupled thermal runaway precursors, lithium plating under fast charging, and mechanical stress from vibration that single-physics monitoring models cannot jointly characterize from onboard sensor streams. Multi-physics digital twins capable of real-time state estimation have not been demonstrated within the CAN bus bandwidth constraints of production fleet IoT architectures. This paper presents DigitalTwinEV, a multi-physics digital twin integrating coupled thermal, electrochemical, and mechanical submodels for real-time battery pack degradation monitoring in fleet EV IoT systems. A reduced-order thermal model solved by a physics-informed neural operator runs at 10 Hz on the vehicle compute unit, predicting cell temperature distributions from current and ambient sensors. An equivalent circuit electrochemical model with online parameter identification tracks state of charge and lithium plating onset indicator. A vibration-induced stress model derived from finite element analysis simulation informs a fatigue accumulation counter. Ensemble Kalman filter fusion integrates the three submodel state estimates with onboard sensor readings. Validated on a 36-vehicle fleet over 18 months of operational data, DigitalTwinEV predicts capacity fade with 0.94% RMSE and detects fast-charge lithium plating events with 91.3% sensitivity at 8-minute lead time. The framework defines multi-physics state estimation as the standard for production fleet battery intelligence.

Fabric defect detection in high-speed industrial weaving requires computer vision systems that generalize across yarn types, weaving patterns, and lighting conditions that vary substantially across production runs, making re-annotation for each product configuration economically prohibitive. Synthetic training data generated from physically accurate textile rendering can reduce real-data annotation cost but requires CFD-informed fluid-structure interaction models to realistically simulate yarn drape and surface deformation. This paper presents TextileDefectCV, a domain-generalized defect detection framework integrating CFD-informed synthetic textile texture generation for industrial weaving IoT quality control systems. A RANS fluid-structure interaction simulation of yarn weaving dynamics generates geometrically accurate fabric surface meshes that are ray-traced under variable illumination to produce photorealistic synthetic defect training images. A domain adversarial convolutional detector trained on mixed real and synthetic data is fine-tuned per production run using a 50-sample real-data adaptation module. Evaluated on the AITEX fabric defect dataset and a 14-class proprietary textile defect corpus collected from Pakistani weaving mills, TextileDefectCV achieves 94.7% mAP, adapts to new fabric types in under 3 minutes, and reduces real-data annotation cost by 83% over fully supervised baselines. The system establishes CFD-physics-grounded synthetic data as a scalable strategy for industrial textile computer vision.

Medical named entity recognition in multilingual electronic health record systems must handle Arabic, Urdu, French, and English clinical notes within unified inference pipelines that standard monolingual BERT variants cannot serve without per-language model instances that multiply GPU memory requirements. Mixture-of-experts architectures offer per-language specialization at shared compute cost but have not been systematically evaluated for cross-lingual clinical NLP. This paper proposes ClinicalMoEBERT, a hierarchical mixture-of-experts clinical BERT model with cross-lingual adapter modules for medical named entity recognition across multilingual EHR systems. A language identification gate routes inputs to language-specialized expert feedforward networks within a shared mBERT backbone, while clinical domain adapters fine-tuned on MIMIC-IV, CLEF eHealth, and Arabic clinical corpora provide domain grounding per language. Cross-lingual entity alignment loss encourages consistent entity boundary detection across languages for the same clinical concept. Evaluated on the i2b2/n2c2, QUAERO French medical NER, and an original Arabic clinical NER dataset from King Abdulaziz Medical City, ClinicalMoEBERT achieves macro-F1 of 88.3% across four languages, outperforming XLM-RoBERTa and mDeBERTa clinical baselines by 6.1 to 9.4 percentage points while operating within a single GPU memory budget. The framework advances multilingual clinical intelligence for global EHR deployments.

Urban water distribution networks lose 20 to 40 percent of treated water to undetected pipe leaks that conventional threshold-based SCADA alarms miss during low-demand nocturnal windows when pressure transient signatures are most diagnostic. Federated learning across municipal water network zones enables collaborative leak model training without centralizing hydraulic topology data considered sensitive national infrastructure. This paper presents WaterNetIoT, an agentic federated learning framework for joint pipe leak detection and pressure anomaly prediction in urban water distribution IoT sensor networks. A graph attention network models pressure sensor inter-dependencies along the pipe network topology, and a federated temporal convolutional network trained across 12 zone-level edge servers learns leak-indicative pressure gradient patterns. An agentic monitoring agent autonomously dispatches inspection work orders through a connected field management API when leak probability exceeds a risk-weighted threshold. Evaluated on the BattLeDIM Battle of the Leak Detection and Isolation Methods dataset and a proprietary sensor dataset from the Lahore Water and Sanitation Authority, WaterNetIoT achieves leak localization within 50 meters in 91.4% of cases, reduces water loss by 23%, and generates actionable inspection dispatch with false dispatch rate of 8.2%. The framework establishes intelligent agentic water loss management for smart urban infrastructure.

Medical LLM inference on shared cloud infrastructure exposes patient query content to side-channel memory attacks and curious cloud provider monitoring that violates HIPAA de-identification requirements even when query data is encrypted in transit. Trusted execution environments protect inference workloads from privileged software attacks but impose latency overheads that exceed acceptable clinical response budgets for large model deployments. This paper presents SecureLLMInfer, a trusted execution environment-protected LLM inference framework integrating output differential privacy for confidential medical query processing on shared cloud GPU infrastructure. An Intel TDX confidential computing enclave hosts quantized INT4 Llama-3-70B inference with attestation-verified model weights, ensuring the cloud provider cannot inspect query content or model activations. A Gaussian output perturbation mechanism calibrated to Renyi differential privacy with epsilon of 3.2 masks individual query influence on generated responses while preserving clinical utility above a validated quality threshold. Secure enclaved batching reduces per-query TEE context switch overhead by 67% over naive enclave-per-request designs. Evaluated on MedQA and a proprietary clinical query dataset, SecureLLMInfer achieves clinical utility score of 4.4 out of 5, query response latency of 1.8 seconds for 1024-token generation, and provides formal epsilon-delta privacy guarantees verified by an independent audit. The framework advances production-grade confidential LLM deployment for regulated medical cloud services.

Patient-specific cerebrovascular flow simulation for intracranial aneurysm rupture risk assessment through finite element Navier-Stokes solvers requires 8 to 24 hours per patient geometry, making prospective surgical planning infeasible within clinical decision windows. Neural operator surrogates trained on population-level CFD data fail to respect patient-specific vessel geometry constraints that govern wall shear stress distributions critical for rupture prediction. This paper proposes NeuroCFDHelm, a physics-constrained Helmholtz-decomposed neural operator for real-time hemodynamic flow simulation in patient-specific cerebrovascular digital twins. A geometry-conditioned Fourier neural operator learns the solution map from vessel boundary conditions and inlet velocity waveforms to velocity-pressure field pairs, with Helmholtz decomposition regularization enforcing divergence-free velocity constraints throughout the vessel lumen. A boundary conforming loss penalizes no-slip wall condition violations at patient-specific vessel surfaces extracted from CTA segmentation. Validated on 1,240 patient-specific aneurysm CFD cases from the @neurIST and AneuRisk databases, NeuroCFDHelm predicts time-averaged wall shear stress with 2.1% mean relative error at 320 times the speed of the OpenFOAM reference solver. The framework enables real-time patient-specific hemodynamic assessment for neurovascular surgical planning and IoT-connected post-intervention monitoring.

Technical debt accumulation in large-scale Java enterprise repositories degrades maintainability at rates that developer review cycles cannot reverse without systematic automated refactoring support. Existing LLM-based refactoring tools lack mechanisms to prioritize refactoring targets by quantified technical debt impact or to ensure that generated refactored code preserves behavioral equivalence verified by existing test suites. This paper presents RAGCodeRef, a retrieval-augmented LLM refactoring framework integrating technical debt quantification for automated software maintenance in large-scale Java repositories. A static analysis pipeline computes SonarQube-derived technical debt metrics per class and method, prioritizing refactoring targets by debt density and coupling impact scores. A retrieval module indexes verified refactoring patterns from curated pull request histories in the top-500 Apache and Spring GitHub repositories, grounding LLM refactoring proposals in evidence-backed transformation templates. An equivalence verification agent executes existing test suites against refactored code and iteratively revises the refactoring until all tests pass or flags unresolvable semantic divergences for developer review. Evaluated on the TechDebt and VulnFix benchmark repositories comprising 2.3 million lines of Java, RAGCodeRef reduces technical debt by 34%, achieves behavioral equivalence in 91.7% of automated refactoring attempts, and completes refactoring cycles 5.8 times faster than unaugmented LLM baselines. The framework operationalizes intelligent evidence-driven software maintenance for industrial Java engineering.

Internet of Vehicles fog computing deployments serving V2X communication for traffic management and autonomous driving are targeted by coordinated Sybil and denial-of-service attacks that exploit the broadcast nature of DSRC and C-V2X channels to inject false traffic advisories. Fog nodes must perform intrusion detection within the 10-millisecond V2X message validity window without centralizing vehicle trajectory data that constitutes private mobility behavioral profiles. This paper presents FogSecIoV, a fog-assisted federated intrusion detection framework integrating attention graph neural networks for coordinated cyberattack mitigation in Internet of Vehicles edge ecosystems. A temporal attention graph network models vehicle message sequence dependencies and inter-vehicle spatial relationships to classify coordinated attack signatures that single-message classifiers miss. Federated model aggregation across fog nodes uses a Krum-based robust aggregation operator resilient to Byzantine fog node updates. A lightweight quantized inference model derived through structured pruning meets the 3-millisecond per-message classification budget on fog server hardware. Evaluated on the VeReMi Extended and ASTM IoV attack datasets, FogSecIoV achieves Sybil attack detection AUC of 0.986, coordinated DDoS detection precision of 94.7%, and sustains detection performance under 30% malicious fog participation. The framework advances production-grade federated security for fog-layer IoV infrastructure.

Urban heat island mitigation through smart city planning requires city-scale computational fluid dynamics simulations of turbulent convective airflow between building clusters that standard RANS k-epsilon closures solve inaccurately under complex building wake interactions without prohibitively fine mesh resolution. Machine learning turbulence closures trained on high-fidelity LES data offer accuracy improvements but have not been integrated with smart city IoT sensor assimilation pipelines for operational planning. This paper proposes ThermalRANS, a data-driven RANS turbulence closure framework integrating physics-informed residual neural network correction for urban heat island CFD simulation in smart city IoT planning systems. A gradient boosted residual network learns additive corrections to the standard k-epsilon Reynolds stress tensor from sparse LES training data on canonical building array geometries. An ensemble Kalman filter assimilates temperature and wind velocity measurements from urban IoT weather station networks to update CFD boundary conditions in near-real-time urban airflow simulations. A Bayesian optimization module explores green infrastructure placement configurations to minimize urban heat island intensity under prevailing wind conditions. Validated against LES reference simulations on the CODASC urban canyon dataset and compared to IoT sensor data from Lahore and Dubai smart city deployments, ThermalRANS reduces velocity prediction error by 34% over standard k-epsilon and produces urban cooling design recommendations within a 40-minute simulation cycle. The framework bridges operational smart city IoT sensing and physics-accurate urban microclimate planning.

Perpetual batteryless biosensor IoT networks for continuous health monitoring rely on ambient energy harvesting from body heat and RF sources that deliver highly intermittent power profiles incompatible with synchronous federated learning communication rounds that assume continuous device availability. Existing energy-aware federated learning schedulers model harvesting as stationary processes that fail under the non-stationary human activity-dependent thermoelectric and piezoelectric generation patterns. This paper presents IoTBioHarvest, an energy harvesting-aware federated learning scheduling framework combining deep reinforcement learning with non-stationary harvesting forecasting for perpetual batteryless biosensor IoT networks. A transformer-based energy arrival predictor models activity-conditioned non-stationary thermoelectric and piezoelectric harvesting processes per device. A proximal policy optimization scheduling agent dynamically selects device participation subsets, compression ratios, and training duration per round to maximize global model convergence under per-device energy buffer constraints. An asynchronous gradient buffering mechanism accumulates partial device updates across multiple harvesting cycles before contributing to aggregation. Evaluated on a custom hardware testbed of 24 batteryless biosensor nodes and the Opportunity activity dataset, IoTBioHarvest achieves training convergence within 94% of the accuracy attained by a battery-powered FedAvg baseline while reducing training-induced energy buffer depletion events by 89%. The framework enables perpetual federated intelligence for energy-autonomous wearable IoT.

Emergency department triage at high-volume hospitals depends on rapid nurse assessments that become inconsistent under staffing pressure, producing undertriage rates for high-acuity presentations that contribute to preventable deterioration. LLM-based triage assistants without retrieval grounding produce acuity classifications that deviate from validated Emergency Severity Index protocols on edge-case presentations. This paper presents MedLLMTriage, a retrieval-augmented LLM triage automation system integrating real-time IoT vital sign streams for high-acuity patient prioritization in smart hospital emergency departments. A vital sign temporal encoder processes continuous SpO2, heart rate, blood pressure, and respiratory rate streams from bedside IoT sensors and retrieves analogous historical triage presentations from a vector-indexed ESI-annotated case library. A chain-of-thought generation module produces structured triage narratives grounded in retrieved analogues and flags high-acuity indicators for immediate nurse review. A conformal set predictor with marginal coverage guarantee outputs ESI level prediction sets that alert on cases where model uncertainty exceeds a calibrated threshold. Evaluated on the MIMIC-IV emergency department dataset and a prospective validation study at a 600-bed tertiary hospital, MedLLMTriage achieves ESI-level agreement with senior nurse triage at 91.8%, undertriage rate of 0.9%, and reduces triage documentation time by 58%. The framework advances safe LLM-augmented triage for IoT-equipped high-volume emergency departments.

Industrial control system digital twins deployed in power generation, water treatment, and oil refinery critical infrastructure require cyber attack attribution capabilities that can identify causal pathways from initial network intrusion to physical process impact, enabling counterfactual resilience planning before adversaries reach operational technology assets. Existing ICS anomaly detectors identify attack presence but cannot attribute causal chains or simulate counterfactual defensive actions. This paper presents CyberTwinICS, a causal attack attribution and counterfactual resilience planning framework for industrial control system digital twins in critical infrastructure IoT environments. A causal discovery algorithm applied to SCADA historian data constructs a process causal graph that an attribution engine traverses backward from detected process anomalies to candidate intrusion entry points. A counterfactual resilience planner simulates alternative network segmentation and patch deployment configurations and predicts their effectiveness in blocking the identified causal attack chains. A risk-ranked remediation report is generated with estimated time-to-implement and operational disruption cost for each recommended countermeasure. Evaluated on the SWaT and BATADAL ICS attack datasets and a custom OPC-UA industrial network emulator, CyberTwinICS achieves attack causal attribution accuracy of 93.2%, counterfactual impact prediction error of 4.1%, and reduces mean remediation planning time by 76% over expert manual analysis. The framework establishes causal intelligence as the foundation for proactive ICS resilience management.

Site reliability engineering for cloud-native microservice platforms requires correlating distributed trace spans, structured logs, and metric time series across hundreds of interdependent services to attribute incident root causes within service level objective recovery windows that manual analysis consistently misses. Natural language processing of incident post-mortem reports encodes organizational failure knowledge that existing automated root cause analysis tools do not exploit. This paper presents NLPDevOps, an NLP-driven incident root cause analysis framework integrating knowledge graph reasoning for automated site reliability engineering in cloud-native microservice platforms. A service dependency knowledge graph constructed from OpenTelemetry trace data and deployment manifests serves as the inference substrate. An LLM fine-tuned on 12,000 incident post-mortem reports from open-source SRE repositories generates structured root cause hypotheses ranked by graph-constrained plausibility scores. A causal graph propagation algorithm traces anomalous metric signals backward through service dependency edges to candidate origin services. A retrieval module grounds hypotheses in analogous historical incidents retrieved from the post-mortem corpus. Evaluated on the DeathStarBench microservice benchmark with injected fault scenarios and a 14-month incident log from a 200-service production platform, NLPDevOps achieves top-3 root cause localization recall of 94.1%, reduces mean time to identify from 38 minutes to 5.7 minutes, and generates audit-grade post-mortem reports rated actionable by SRE teams in 88% of cases.