ABSTRACT. This paper presents a biomechanically grounded simulation framework that generates synthetic, high-fidelity foot-mounted IMU signals with controllable sensor errors. The proposed pipeline uses predictive forward-dynamics musculoskeletal simulations to obtain realistic foot kinematics, which are then processed with smoothing splines and analytical differentiation, and converted via a strapdown inertial formulation combined with a deterministic and stochastic noise model. The framework is validated by comparing synthetic data with physical IMU measurements in straight-line walking and stair-climbing scenarios. Results show that the 3D musculoskeletal simulation outperforms traditional reduced-order models (e.g., the 2-DOF inverted pendulum) in predicting horizontal navigation errors (RMSE of 0.86~m vs. 0.58~m, compared to a 0.94~m experimental baseline) and can also simulate complex, non-level gait dynamics. By capturing the biologically driven micro-structure of human locomotion, the framework offers a controllable and physically grounded tool for robust PDR algorithm evaluation.

ABSTRACT. Compound hydrological events and the proliferation of autonomous IoT-based environmental monitoring initiatives across European public administrations are producing a fragmented information landscape, while the granularity of regional alerting chains does not by itself satisfy inter-municipal civil-protection planning requirements. This paper presents a federated LoRaWAN testbed for hydro-meteorological monitoring deployed at provincial scale across the Province of Rimini, Italy, as a Test Before Invest service of the European Digital Innovation Hub ER2Digit. The deployment comprises twelve ultrasonic hydrometers and eight meteorological stations distributed along the principal watercourses of the province, on a three-tier architecture combining a 1+1 redundant LoRaWAN access layer, the regional IoT platform RetePAIoT as a single structured collector, and a role-based consumption tier preserving the operational autonomy of each participating entity. The testbed is commissioned by the inter-municipal alerting authority of the province, acting on behalf of four municipalities and one Union of Municipalities, and instantiates a federation pattern empirically supported by two precursor deployments in the same regional ecosystem. The deployment is positioned as a knowledge-enrichment layer complementary to the official alerting chain, providing the data plane for forthcoming AI- and digital-twin-based public-safety applications

ABSTRACT. The National Institute of Standards and Technology’s Public Safety Communications Research Division (PSCR) engages with public safety partners to help address their urgent needs. In these conversations, public safety frequently highlights the need to monitor personnel vertical location in a building, or their position along the z-axis. Toward enabling accurate tracking technologies in three dimensions, we developed a location tracking testbed that focuses on accelerating the pace of technological development in the fields of indoor mapping, tracking, and navigation for public safety stakeholders. This testbed consists of 50 motion-capture cameras capable of continuous tracking with sub-centimeter accuracy in three dimensions, including the z-axis. This system is located within a multistory building with approximately 15 m in the vertical axis and provides high-frequency capturing. This z-axis testbed also contains ground-truth reference points surveyed by the National Geodetic Survey (NGS). Combining these reference points with the motion capture system, this testbed will be able to: i) evaluate the systems vertical and horizontal positioning accuracy in multi-floor tracking systems, ii) benchmark sensor-based localization approaches including inertial, radio frequency, and vision-based methods, iii) model vertical and horizontal error propagation and calibration needs in 3D indoor positioning, and iv) create a test environment for infrastructure-free indoor localization, mapping, and navigation solutions. This paper discusses the technical development of the z-axis testbed in Boulder, CO, and presents initial tracking capabilities using a cell phone and passive tracking markers.

ABSTRACT. Deploying broadband Mission-Critical Push-To-Talk (MCPTT) services over shared commercial infrastructures introduces resource contention during multi-agency responses in mass-crowd events. This study evaluates cross-carrier interoperability and standard versus prioritized quality of service (QoS) frameworks under real-world saturation constraints. We design an empirical multi-carrier field experiment utilizing twelve identical smartphones deployed across multiple physical sectors inside Texas A&M University's Kyle Field during a football game with 105,000+ attendees. Automated voice calls were monitored using Perceptual Objective Listening Quality Analysis (POLQA), packet delivery metrics, and connection rates. The results reveal that voice path failure is isolated to network infrastructure bottlenecks rather than device hardware limitations. Specifically, we identify a sharp, non-linear network failure model where transport-layer jitter exceeding a critical threshold de-jitter buffer underflows, causing structural audio degradation. Priority-managed channels effectively bypass this congestion. This study helps establish an operational insight for emergency planners to mandate network infrastructure, end-to-end network slicing and dedicated resource provisioning capable of keeping transport-layer jitter below the critical failure boundary.

ABSTRACT. Public safety and physical security operations rely on multiple independent systems for alarm monitoring, personnel tracking, and incident reporting, which often operate in isolation and limit operational visibility. They are typically a set of separate systems which prevent one from reconstructing and assessing the association of sensor-activated events to human response. We present a common geospatial decision-support framework based on IoT-enabled alarm data, guard location tracking, and incident workflows into an operational platform.

We propose a model to combine deterministic geospatial calculation via PostgreSQL/PostGIS with a hybrid interaction layer based on large language model (LLM) to interpret queries and explain results. The structured query architecture separates operational, exploratory, and interpretive interactions, allowing for a safe and reliable use of LLM capabilities. Moreover, it introduces a quantitative score model evaluating coverage of events, response effectiveness, and spatiotemporal activity density. This framework enables real-time and retrospective visualization of situational awareness and operational metrics. This integration allows the system to effectively create a dynamic Common Operational Picture, as it connects sensor-generated events tightly to personnel response workflows to facilitate the visibility, evaluation, resource utilization, and monitoring that is vital in the safety of the public as a whole.

ABSTRACT. Abstract—Public safety command and control centers increasingly depend on heterogeneous technologies spanning field sensing devices, surveillance systems, communications platforms, physical security controls, analytics engines, and geospatial operating environments. In many deployments, however, these capabilities remain fragmented across proprietary subsystems, limiting interoperability, scalability, and coordinated crisis response. This paper proposes a vendor-neutral reference architecture for AI-enabled public safety command and control centers informed by retrofit and greenfield smart city deployments. The architecture is organized into seven logical layers: field device, edge analytic gateway, information management, analytics, presentation, command-center-to-command-center integration, and cross-cutting governance. A key contribution is the use of edge analytics to augment field intelligence, particularly where legacy devices can be enhanced without full hardware replacement. The architecture also introduces a generalized emergency operations data pipeline for integrating operational sources through API, streaming, file-based, and geospatial ingestion into normalized, analytics-ready data products. At the presentation layer, AI-enabled insights are combined with map-first dashboards, geospatial views, digital twins, and automated situation reporting to support both reactive and predictive incident management. Validation across two smart city contexts indicates that a layered vendor-neutral approach improves adaptability, multiagency coordination, and resilience while reducing dependence on monolithic command platforms.

ABSTRACT. Abstract— Gas leak detection and gas environment classification are critical for industrial, residential, environmental, and public safety applications, especially in situations where first responders and emergency management teams need early information about hazardous gas conditions. Low-cost gas sensors are widely used for gas monitoring; however, sensor-only detection can be limited in complex environments involving mixed gases, low concentrations, and noise. Thermal imaging provides an additional modality by capturing temperature-related patterns in the surrounding environment. This study presents a three-stage artificial intelligence framework for gas environment classification using the Multimodal Gas dataset. The dataset contains synchronized measurements from seven MQ gas sensors and thermal camera images for four classes: NoGas, Perfume, Smoke, and Mixture. In the first stage, several traditional machine learning models are trained and compared using only MQ sensor readings. In the second stage, a convolutional neural network is developed to classify gas environments using thermal images only. In the third stage, a multimodal fusion model combines CNN-based image features with neural-network-based sensor features for final classification. The models are evaluated using accuracy, macro precision, macro recall, macro F1-score, and confusion matrices. The results show that multimodal fusion provides the most effective classification performance compared with sensor-only and image-only approaches. This staged analysis demonstrates the potential of AI-assisted gas sensing systems for complex environments and provides a foundation for future field-deployable UAV-based gas monitoring systems using Long Period Fiber Grating (LPFG) sensors to support first responders and emergency management operations.

ABSTRACT. Indoor environments lack the spatial intelligence infrastructure that GPS provides outdoors; first responders ar- riving at unfamiliar buildings typically have no machine-readable map of safety equipment. Prior work on 3D semantic segmenta- tion for public safety identified two barriers: scarcity of labeled indoor training data and poor recognition of small safety-critical features by native point-cloud methods. This paper presents INSIGHT, a zero-target-domain-annotation pipeline that projects 2D image understanding into 3D metric space via registered RGB-D data. Two interchangeable vision stacks share a common 3D back end: a SAM3 foundation-model stack for text-prompted segmentation, and a traditional CV stack (open-set detection, VQA, OCR) whose intermediate outputs are independently inspectable. Evaluated on all seven subareas of Stanford 2D- 3D-S (70,496 images), the pipeline produces Pointcept-schema- compatible labeled point clouds and ISO 19164-compliant scene graphs with ∼104 × compression; role-filtered payloads transmit in <15 s at 1 Mbps over FirstNet Band 14. We report per-point labeling accuracy on 7 shared classes, detection sensitivity for 15 safety-critical classes absent from public 3D benchmarks alongside code-capped deployable estimates, and inter-pipeline complementarity, demonstrating that 2D-to-3D semantic transfer addresses the labeled-data bottleneck while scene graphs provide building intelligence compact enough for field deployment.

ABSTRACT. Online adaptive radiotherapy (oART) is a mission-critical clinical technology in which AI-generated contours, rapid plan adaptation, and real-time delivery decisions must operate reliably under severe time pressure. These conditions mirror broader public-safety environments, where cascading failures, human–AI coordination gaps, and temporal complacency can compromise operational integrity. To characterize these risks, we develop a Monte Carlo simulation framework modeling failure propagation across a nine-stage adaptive workflow with 47 parameterized failure modes and interdependent escalation pathways. Using this model, we evaluate the Risk Assessment and Dimensional Adaptive Reliability (RADAR) architecture multilayer assurance framework designed to strengthen resilience in safety-critical AI systems. Across 100,000 simulated fractions, full RADAR reduces overall error rates by 18.6% and severe events by 29.2%, increasing conditional interception from 0% to 19.5%. Layer ablation analysis shows that Risk Assessment and Adaptive Culture account for 82.6% of all prevented failures, underscoring the importance of organizational vigilance and structured human–machine cross-checks. Modality analysis reveals that CBCT (cone-beam computed tomography)-guided workflows exhibit 52.5% higher error rates than MR-guided workflows (16.94% vs. 11.11%), with MR achieving a substantially higher proportion of zero-error treatment courses. These results position RADAR as a scalable public-safety-oriented framework for improving reliability, resilience, and risk governance in AI-enabled clinical operations.

ABSTRACT. Emergency communication centers are transitioning to Next-Generation 911 (NG-911) infrastructure supporting multimodal emergency reporting through voice, text, images, and video. However, their current dispatcher training processes predominantly rely on role-playing, requiring trainees to develop diverse cognitive and functional skills, such as semantic interpretation of diverse calls across NG-911’s multimodal communication context supporting voice, text, images, and video. Traditional role-play methods are resource-intensive, requiring experienced instructors to manually enact caller scenarios, limiting scalability, and thus, inadequately addressing evolving multimodal communication requirements. This paper introduces a novel system, MATE-911: Multimodal AI-powered Training Environment for Next-Generation 911 Dispatcher Preparedness, an AI-driven platform filling the NG-911 training gap through automation of multimodal emergency call scenario generation. MATE-911 uses Retrieval-Augmented Generation (RAG) approach to ground content generation in validated knowledge of emergency response protocols, thus avoiding hallucinations, which are crucial in safety training domains. Ontology-based validation of the generated content according to official NENA protocols is incorporated into the process so that the generated content adheres to dispatcher standards before being presented to the trainees. A MATE-911 pilot was developed using open-source tools and modeling frameworks to enable local deployment without dependency on commercial APIs. Specifically, we employed Llama 3.1 8B for language model generation and ChromaDB for vector database storage, allowing agencies to deploy the system on commodity hardware without archived call databases or commercial API dependency helping adaptation in resource limited organizations. The current implementation fully validates text and voice modalities, image and video generation are architecturally designed and designated for future implementation. Technical assessment using 160 gold-standard scenarios achieved a protocol F1 score of 0.81 with a 3.0% hallucination rate and sub-2-second generation latency on commodity hardware, indicating a promising application of AI for developing scalable, protocol-compliant NG-911 dispatcher training tools.

ABSTRACT. Conservation of scarce spectrum allocated to 5G public safety bands depends on adaptive and efficient Modulation and Coding Scheme (MCS) selection algorithms. Sidelink Mode 2, used for off-network operation, differs from infrastructure modes with respect to signaling, distributed scheduling, and bandwidth granularity. In this setting, we evaluate how well the popular Outer Loop Link Adaptation (OLLA) algorithm performs, and compare it with a reinforcement learning approach and with an idealized (oracle) controller. We find that basic OLLA operates close to the oracle’s performance in static environments, but that its resource footprint can be minimized by jointly selecting the MCS and the number of retransmissions based on estimates of channel conditions. We also analyze the performance of OLLA in a busy channel and demonstrate that OLLA is unable to distinguish whether negative acknowledgments are caused by interference, half-duplex events, or channel conditions. This limitation is analogous to longstanding difficulties with TCP congestion control in wireless networks and warrants further re- search on OLLA extensions and other adaptive MCS algorithms for the sidelink Mode 2 environment.

ABSTRACT. Autonomous systems rely on probabilistic state estimation to support safety-critical decision-making. However, under adversarial sensor manipulation, these estimators may remain statistically consistent while diverging from physical reality, creating a failure mode in which unsafe actions are executed with high confidence. In this work, we demonstrate this phenomenon in an evaluated stealth-drift GNSS spoofing scenario, where an Extended Kalman Filter (EKF) maintains low estimation uncertainty while producing incorrect control decisions that lead to stop-line violations. To address this limitation, we propose Causal Reality Anchoring (CRA), a decision-level validation framework that enforces physical consistency across independent sensing modalities before control execution. CRA introduces constraint-based validation to detect persistent inconsistencies between spoofed GNSS measurements and physically grounded sensors such as lidar, enabling unsafe candidate actions to be rejected before execution. Monte Carlo simulations under the evaluated stealth-drift GNSS spoofing scenario show that CRA prevents unsafe stop-line violations in all tested runs, reducing the violation rate from 100% for the EKF-only baseline to 0%, while maintaining zero false positives under nominal conditions. These results do not imply universal protection against all GNSS spoofing attacks; rather, they demonstrate that estimation consistency alone may be insufficient for safety under stealthy adversarial drift and highlight the importance of physically grounded decision-level validation in autonomous systems. Index Terms—Autonomous systems, GNSS spoofing, sensor fusion, Extended Kalman Filter, safety validation, multi-sensor consistency, decision-level verification, lidar-GNSS fusion

ABSTRACT. Transportation is responsible for more than 16% of greenhouse emissions with 99% of the world's urban population breathing polluted air. Active travel is a promising alternative to motorized travel that aims to limit the negative impact of transport on air quality. However, people who adopt active travel (e.g., walking or cycling) are considered vulnerable road users (VRU) and are at risk of fatal accidents as they share the road with vehicles. Intelligent transportation systems (ITS) augment transport infrastructures and vehicles with information and communication technologies to limit emissions while improving road safety. In this work, we examine emerging enabling technologies for ITS and their applications for improving air quality and VRU safety. To this end, we investigate distributed acoustic sensor (DAS) technology as an alternative method for sensing mobility and we propose a deep-learning model for interpreting DAS complex data. The safety of two types of VRUs: pedestrians and cyclists, is next studied using camera-based sensors and ensemble learning methods. Data safety in ITS is critical as it relates to sensitive and private information. This work examines federated learning and blockchain as effective technologies to ensure data safety whilst allowing the applications of sensors and machine learning.

ABSTRACT. Autonomous vehicles (AVs) operating in urban environments have been documented to interfere with emergency first-responder operations through failure modes including freezing, traffic-law violations, and non-compliance with officer hand signals. This paper presents ERAV-Q (Emergency-Responsive Autonomous Vehicle framework with Quantum optimization), a simulation-validated AI framework addressing these failure modes through three integrated components: (1) a real-time emergency vehicle detection module using a fine-tuned YOLOv8-based perception pipeline augmented with V2X communication, (2) an LLM-based gesture recognition system leveraging a vision- language model for interpreting first-responder hand signals under diverse environmental conditions, and (3) a Quadratic Unconstrained Binary Optimization (QUBO) solver for quantum-assisted AV priority routing in multi-vehicle corridors. Simulated evaluations in a CARLA-based urban environment demonstrate that ERAV-Q reduces AV-induced emergency vehicle delay by 68.4% over a baseline rule-based system, achieves 91.3% gesture classification accuracy across 14 standardized traffic-control gesture classes, and produces conflict-free priority routes for up to 12 simultaneous AV–emergency vehicle interactions with a mean solve latency of 47 ms on a D-Wave hybrid sampler. These results suggest that targeted AI intervention at the perception, interpretation, and planning layers can materially reduce the operational burden imposed on emergency responders by deployed AV fleets.

ABSTRACT. The operational integrity of autonomous mobile robots relies on the continuous availability of wireless control loops, making them highly attractive targets for adversarial intentional electromagnetic interference. This paper introduces a resilient, cross-layer framework that combines physical-layer spectral perception with network-layer routing optimization to protect middleware stability, such as ROS~2, during intentional electromagnetic interference. Utilizing a software-defined radio front-end, the system extracts dynamic spectral descriptors, including spectral entropy and channel occupancy, to inform a Random Forest classifier that establishes adaptive environmental baselines. To ensure uninterrupted data flow, the architecture maintains dual pre-authenticated physical interfaces in a hot-standby configuration, enabling instantaneous failover through automated network routing table updates. Empirical validation on a physical ROS~2 mobile robot testbed demonstrates that this adaptive hardware-assisted architecture optimizes communication recovery to an average of 141~ms. This sub-second restoration translates directly into a 78.9\% reduction in pooled root-mean-square path tracking error compared to software re-association, successfully securing system-level mission integrity.

ABSTRACT. This paper presents a spatio-temporal GNSS resilience framework is introduced to combat multipath, signal blockage, spoofing, and jamming in a multiconstellation scenario. The proposed framework integrates multi-constellation data acquisition from NASA’s CDDIS data repository, ionospheric pseudorange correction, and trilateration-based positioning, along with a hierarchical ML pipeline that includes a Random Forest spatial classifier and an LSTM-based temporal model. The validation of Phase-I of the pipeline yields a data validity of 96.4%. The hybrid model shows 99.8% accuracy in detecting anomalies in Normal, Multipath, Spoofing, and Jamming categories.

ABSTRACT. Future 5G-V2X and beyond-5G vehicular systems must sustain reliable low-latency communication under high mobility, dynamic blockage, beam instability, and heterogeneous radio access conditions. Existing networkcentric control approaches mainly react to degraded radio key performance indicators (KPIs), which may delay response to fast physical-world changes affecting mmWave links. Although recent sensing-aided wireless studies show that multimodal perception can support beam and blockage-related prediction, most approaches stop at prediction outputs and do not translate them into O-RAN-consumable control information. This paper proposes a multimodal BEV representation and spatialtemporal framework for 5G-V2X O-RAN control and connectivity management. The framework constructs a blockage-aware multimodal Bird’s Eye View (BEV) representation by enriching LiDAR BEV geometry with UE position and RSU-to-UE line-of-sight (LOS) context, fuses it with temporal radio features and LOS-aware camera semantic cues, and converts the learned state into O-RAN risk tuples containing candidate beam lists, uncertainty, communicationblockage probability, and policy primitives. Using DeepSense 6G Scenario 31, the multi-task BEV-radio-camera semantic LOS model achieves 55.99% Top-1, 89.00% Top-3, and 96.10% Top-5 beam prediction accuracy under communicationblockage supervision, showing modest improvement for blockage-aware beam-list preparation. More importantly, the framework translates beam and blockage predictions into structured O-RAN risk tuples containing candidate beam lists, communication-blockage and link-risk probabilities, beam uncertainty, and policy primitives such as beamList, enableDC, preferRAT, controlUrgency, sliceAssist, and recommended xApp action. The blockage-aware tuple policy achieves 84.58% F1-score, 74.00% precision, 98.67% recall, and 1.33% missedrisk rate against the O-RAN control-risk proxy. These results demonstrate that the proposed approach moves beyond prediction-only V2X intelligence toward an O-RAN-ready multimodal state representation for proactive beam and connectivity management. The high safety-oriented triggering behavior also indicates that practical near-RT RIC deployment requires xApp guides such as live KPI confirmation to reduce unnecessary proactive actions.

ABSTRACT. Introduces a digital twin -driven approach for assessing electrical busways safety performance. An operational-based parameter set “trains” machine learning models for anomaly safety detection, including useful life estimation. Publication overviews UL857/IEC61439-6(Busway/Fitting) influenced fault identification, degradation analysis, and proactive maintenance of power distribution systems.

ABSTRACT. Today’s emergency response systems are based on disjoint approaches where situational data, resource inventories and decision-support functions are latched in isolated silos, causing significant life-saving delays. This paper introduces the Adaptive Digital Twin Emergency Coordination (ADTEC) framework, a multi-layer hierarchical architecture, harnessing AI-driven digital twin analytics and real-time data from IoT sensors to facilitate proactive and coordinated emergency response across diverse urban environments. The first contribution is ADTEC’s Hierarchical Federated Learning (HFL) module which allows for privacy-preserving, inter-agency model training without centralizing sensitive operation data. The second is a Quantum-Assisted Optimization Engine (QAOE) that decreases the computation time for multi-depot vehicle routing by 73.2% in simulation benchmarks over classical metaheuristic baselines, and the third is a Temporal Graph Neural Network (TGNN)-based anomaly detection layer capable of identifying emerging multi-hazard cascades with a lead-time of 8-14 minutes before human analysts. Benefits of the product are shown by reducing average deployment latency of first responders by 34.7%, underutilization of resources by 29.1%, and covering more people when the size of the first responders’ fleet is limited by 41.3% across three archetypal emergency scenarios: earthquake response in the urban domain, multi-building fire escalation, and coordinated mass-casualty incidents. The modularity of the framework satisfies NIST SP 800-160 security requirements, letting ADTEC be a deployable and standards-compliant solution to smart city emergency management infrastructure.

ABSTRACT. Public safety systems require rapid, coordinated decision-making in highly dynamic and uncertain environments. Traditional centralized architectures struggle to meet these demands due to latency constraints, limited scalability, and the absence of real-time feedback mechanisms. Recent advances in artificial intelligence, edge computing, and digital twin technologies offer promising capabilities; however, existing solutions often treat prediction, simulation, and decision-making as disjoint processes.

This paper proposes a unified, closed-loop framework that integrates digital twins with multi-agent reinforcement learning (MARL) under a centralized training, decentralized execution (CTDE) paradigm for real-time public safety optimization. The proposed system introduces a state synchronization mechanism that continuously aligns physical and virtual environments, enabling adaptive and feedback-driven decision-making. Edge-based semantic compression reduces communication overhead and latency, while the digital twin provides a global state representation for coordinated multi-agent control.

The framework is evaluated using a high-fidelity urban simulation with heterogeneous agents representing emergency response units. Experimental results demonstrate a 38.7% reduction in response time, over 92% prediction accuracy, and significant improvements in resource utilization and system latency. Ablation studies confirm that the digital twin component plays a critical role in enhancing coordination and stability.

The proposed approach establishes a scalable and practical architecture for next-generation intelligent public safety systems, bridging the gap between predictive analytics and real-time operational decision-making.

ABSTRACT. Nowadays, urban public safety depends on coordination between human activity and transport systems during major events, such as crowd surges or emergency evacuations. However, existing intelligent transportation systems (ITS) and Vehicle-to-Everything (V2X) frameworks treat buildings as passive structures rather than as contributors to decision-making for outdoor traffic conditions. We introduce Building-to-Everything (B2X), which enables smart buildings to serve as active participants collaborating with urban traffic control. Here in B2X, Building Edge Nodes (BENs) process sensor data from real smartwalls we built and generate structured safety events. Then, these events are transmitted through B2X so that traffic control systems can use them to make proper decisions during any emergencies and large crowds. We built a prototype integrated with the SUMO traffic simulator to evaluate performance during stadium surges and along evacuation routes. Additionally, B2X also improves traffic control with earlier response, reducing congestion and ensuring public safety. With B2X, we can build a smart city where buildings can make decisions in emergency situations by collaborating with nearby smart buildings, traffic signals, transportation systems, and other systems.

ABSTRACT. Public safety agencies increasingly deploy mission-critical systems---911 dispatch platforms, emergency alert systems, first responder data networks---on Azure cloud infrastructure. These systems are provisioned through Infrastructure as Code (IaC) templates written in Azure Bicep. A single misconfiguration in a Bicep template can expose sensitive incident records to the public internet, permit unencrypted data transmission over emergency communication channels, or cause deployment failures that take critical public safety systems offline. This paper presents BICEPTOOLING, a seven-pass multipass static analysis architecture for Azure Bicep, implemented in C\#~(.NET~10). The system detects security misconfigurations before deployment (rules SEC001--SEC010), provides guided error diagnostics (BCP001--BCP006) that teach developers the violated Azure rule at the moment of the mistake, and includes a Bicep Explainer pass producing plain-English descriptions of cloud infrastructure declarations. An interactive ten-lesson learning platform trains the developers who build public safety cloud systems from zero experience to correct deployment. The prototype achieves 100% pass rate across 14 unit tests, correctly transpiles Bicep to deployable ARM~JSON, and analyzes 493 of 500 real-world Azure templates (98.6%), detecting misconfigurations in 43.6% (SEC004), 26.6% (SEC008), and 13% (SEC001/SEC002) of templates---vulnerabilities directly relevant to CJIS and FIPS~140-2 public safety data protection requirements. A CodeBERT multi-label classifier trained on the labeled corpus achieves AUC-ROC 0.774 (SEC009) and 0.752 (SEC008). A Random Forest trained on 36 rule-specific structural features achieves macro AUC-ROC~0.984 across 16 rules (+0.362 over CodeBERT), with all 16 trainable rules exceeding AUC-ROC~0.75.

ABSTRACT. As hospitals become more digital, systems such as electronic health records, radiology platforms, laboratory services, medication systems, and networked medical devices are becoming increasingly interconnected and complex. This growing complexity increases the risk of cyber threats that can disrupt hospital operations and affect patient safety. To address this challenge, we introduce TwinGuard, a digital twin framework designed to evaluate the cyber and public safety risks of newly hired hospital employees before their accounts are activated on the live network. Rather than relying only on manual review or fixed role templates, TwinGuard allows hospital administrators and security teams to simulate the possible impact of granting specific access privileges. Considering factors such as job role, requested systems, missing security controls, and background check uncertainty, the framework analyzes how an account could reach critical hospital systems, including EHR, PACS, pharmacy, and device management services. TwinGuard combines user centric digital twins, graph based attack path analysis, weighted risk scoring, and rule based decision logic to determine whether access requests should be approved, restricted, reviewed, or rejected. The system helps identify excessive permissions, detect risky attack paths, and recommend controls before access is granted.