ABSTRACT. Autonomous drones and connected vehicles increasingly rely on swarm-level cooperation for collective sensing, navigation, and decision-making, making them vulnerable to coordinated attacks that traditional per-device security mechanisms cannot effectively address. This paper presents a decentralized swarm-level security framework that combines collective anomaly detection, trust-aware consensus, majority-vote isolation, DoS filtering, and self-healing recovery to defend against Byzantine, Sybil, DoS, and stealthy attacks. The framework leverages peer cross-validation and asymmetric trust memory to suppress malicious influence while preserving resilient swarm operation under adversarial conditions. Experimental results demonstrate improved attack resilience, bounded consensus degradation, scalable near-linear performance in sparse topologies, and effective recovery from false-positive isolations, highlighting the importance of fleet-centric security for trustworthy deployment of cooperative autonomous swarms.

ABSTRACT. Multi-agent coordination in public safety applications increasingly relies on decentralized information fusion, where agents share and aggregate local estimates for real-time collective decision-making. Trust-aware fusion improves robustness by assigning adaptive credibility weights based on estimated agent reliability, but its behavior under coordinated adversarial manipulation remains insufficiently understood. This paper evaluates the robustness limits of trust-aware fusion in multi-agent search-and-rescue estimation scenarios affected by sensing noise, communication degradation, and adversarial spoofing. Through Monte Carlo simulations, we compare mean-based, median-based, and trust-aware fusion as adversarial penetration increases. The results show that trust-aware fusion can reduce error relative to naive averaging when adversarial influence is limited. However, its performance degrades when coordinated adversaries corrupt the consensus structure, causing biased agents to appear mutually credible while truthful agents may be penalized as inconsistent. Median-based fusion remains more stable under stronger adversarial influence because it is less dependent on consensus-derived trust scores. These findings highlight a limitation of consensus-dependent trust mechanisms and motivate hybrid fusion strategies that combine adaptive trust estimation with robust statistical aggregation.

ABSTRACT. We present a real-time system for detecting threatening approach behavior using egocentric vision, focusing on close-proximity confrontation scenarios. Our system produces a feature vector of pose detection keypoints capturing pose skeleton geometry, camera-compensated optical flow, and motion dynamics weighted by proximity. A three-stage optical flow decomposition separates subject radial expansion from background motion caused by the camera, compensating for the camera movement inherent in the egocentric viewpoint footage. Three lightweight neural networks, trained on a large number of naturalistic videos, have been evaluated on two independent test sets: a within-domain test set with a subject different from training and a cross-domain test set. The system is deployed on a Raspberry Pi 5 with a Hailo-8 NPU at 10 Hz, demonstrating its feasibility for real-world use without dedicated server infrastructure. We make our code publicly available at: https://github.com/arw5902/live-assault-detection.

ABSTRACT. Connected autonomous vehicles increasingly rely on information exchanged through V2X communications to support safe and efficient decision-making. However, the presence of unreliable or malicious information providers creates significant challenges for trustworthy information acquisition. Existing bidding and auction mechanisms primarily focus on economic optimization, while many trust-management approaches do not jointly consider trustworthiness and acquisition cost during bidder selection. To address this limitation, this paper proposes a Belief Rule-Based (BRB) trust-aware bidding framework that integrates provenance, reputation, and consistency evidence to evaluate the trustworthiness of information providers under uncertainty. The resulting trust score is combined with bid price through a utility-based optimization mechanism to support secure and cost-effective information acquisition. The framework is evaluated using trust attributes derived from VeReMi vehicular communication traces and compared with reputation-based, threshold-based, and weighted trust approaches. Experimental results demonstrate that the proposed BRB framework provides more effective identification of trustworthy information providers while reducing the influence of malicious participants and maintaining strong economic performance. These findings highlight the potential of provenance-aware BRB reasoning for secure information marketplaces in connected autonomous vehicle environments.

ABSTRACT. Covert wireless communication seeks to maintain reliable transmission between legitimate parties while preventing a passive warden from detecting radio frequency emissions. By ensuring reliable transmission while preventing adversarial detection without shared keys, this framework addresses critical public safety technology needs for secure, resilient tactical communications in hostile or sensitive environments. This paper investigates keyless spatial covert communication under a practical energy-detection threat model, eliminating the requirement for complex channel state information feedback. Instead, we model a directional spatial masking strategy using geometric spatial shaping combined with artificial noise injection to degrade an adversary's (Willie) detection capability while securing a reliable link to an intended receiver (Bob). Using a unified framework of simulations and software-defined radio testbed experiments, we evaluate the reliability-covertness trade-off across varying physical configurations and power allocations. Scaling the antenna layout concentrates energy toward Bob, creating an angle-dependent detection landscape where Willie’s detection probability approaches baseline false-alarm rates off-axis while Bob's bit error rate stabilizes below $10^{-3}$. Conversely, experimental validation exposes critical physical-layer boundaries: short-range, low path-loss regimes introduce performance ceilings due to persistent ambient energy leakage. These findings reveal a fundamental gap between idealized spatial abstractions and physical constraints, offering key design insights for environment-aware covert physical layer security

ABSTRACT. The growing applications of Large Language Models (LLMs) across public safety and healthcare fields — from emergency triage to patient-tracking — has introduced serious risks of sensitive medical data leakage, particularly when handling multi-modal data such as clinical text, medical images, and physiological signals. Current privacy-protecting approaches are limited because they are often applied in isolation and lack continuous monitoring. To address these limitations, we propose a unified multi-modal Privacy-Enhancing Framework (PEF) that integrates data extraction, modality-specific protection, and real-time monitoring into a single pipeline. PEF is demonstrated on a structured clinical admissions dataset, with evaluation focused on PHI detection, masking effectiveness, and scalability. Experimental results show reliable identification and protection of sensitive information while preserving the structure needed for downstream analysis. These findings indicate that PEF is a viable privacy-enhancing technology for the secure use of LLMs in public safety and healthcare settings.

ABSTRACT. Artificial intelligence (AI) has become a more common choice for workforce-related decision making, such as incidents of talent acquisition, employee retention forecasting, evaluating employee performances, and workforce planning. Nevertheless, the ubiquity of black-box models has raised concerns among enterprise stakeholders about issues of transparency, trust, fairness and regulatory compliance. This work studies the use of Explainable Artificial Intelligence (XAI) to reduce these challenges in workforce prediction systems. An explainable machine learning framework is designed based on SHAP and local interpretability techniques to interpret the data regarding workforce in real life as well as feature-level interpretability. Experimental results show that explainability is an important aspect by which to measure expert trustworthiness of model outputs: in experiments trust in model outputs raised by over 35% compared to non-explainable approaches. In addition, the framework allows for effective human-ai interaction by allowing domain experts to analyze and validate the predictive results. These findings underpin the importance of explainability as a primary requirement for the responsible deployment of AI-driven workforce analytics applications within the enterprise environment.

ABSTRACT. Wildfire evacuation at the wildland–urban inter- face (WUI) fails at the intersection of two coupled problems: computing effective routes under a time-varying fire front, and communicating those routes to threatened populations before the fire reaches them. Recent disasters—the 2018 Camp Fire (85 fa- talities), the 2018 Woolsey Fire (295,000 evacuees, widespread gridlock), and the 2025 Eaton Fire (alerts delayed more than two hours in the most heavily impacted neighborhoods)—show that routing and alert dissemination cannot be treated as independent problems. This paper presents a system that pairs a quantum- enabled Vehicle Routing Problem (VRP) solver, formulated as a Quadratic Unconstrained Binary Optimization (QUBO) Hamilto- nian solved with a Variational Quantum Eigensolver (VQE), with a solver-driven phased alert dispatch layer that maps directly to existing Integrated Public Alert and Warning System (IPAWS), Wireless Emergency Alert (WEA), and Emergency Operations Center (EOC) infrastructure. The solver accepts input from any upstream fire-spread model through a standardized interface and produces an urgency-prioritized evacuation schedule that drives phased communication to residents and coordinating authorities. Pilot studies on the Camp Fire and Woolsey Fire geographies demonstrate 55% and 63% reductions in total evacuation distance, 1.9× and 2.3× faster solve times, and shelter utilization within operational bounds, with dynamic re- optimization evacuating 85% of cells before fire expansion. The system is designed as a decision-support tool for authorized emergency management personnel, using national standards to enable deployment beyond California.

ABSTRACT. Disasters provide images of low quality in the air due to the distortion in the atmosphere, motion blur, debris and smoke, which do not allow manual interpretation and slow the operations in the emergency response. The suggested system integrates a 5G-enabled camera mounted on an unmanned aerial vehicle to capture and transmit high-resolution aerial imagery in real-time to a Multi-Access Edge Computing server for low-latency processing. The received images are processed using a DeepLabV3+ segmentation model that detects damaged regions across twelve distinct classes including buildings, vegetation, roads, water bodies, and multiple levels of structural damage. The segmented output is then analyzed by an EfficientNet-B0 convolutional neural network which quantifies the extent of damage and estimates severity on a five-point scale. The system calculates an impact score through a combination of a predicted severity level and a calculated damage spread ratio obtained through the segmentation mask. The severity classifier has a high accuracy of 95.96% during training and 87.40% during testing, showing promising results in distinguishing between low, medium, and high damage levels. The outputs from each model-such as segmentation masks, severity levels, impact scores, damage percentages, and road blockage detection-are displayed on a web-based dashboard using graphical charts. An API provides a structured summary of the disaster for easy interpretation. This can be extremely helpful to rescue and safety teams to make effective decisions in a timely manner.

ABSTRACT. Adverse weather conditions can significantly affect road safety by reducing visibility, altering vehicle handling, and increasing the risk of accidents. This work proposes a machine learning (ML) approach for predicting the driving safety index (DSI) for highway use. Daily weather data for the city of Toronto, Canada was used to train two ML models - Categorical Boosting and Random Forest to predict driving related weather conditions. Weighted geometric mean was then used to aggregate the resulting predictions and to compute the DSI. Obtained results were benchmarked against real world data from the Toronto traffic collision dataset and two weather prediction websites, and were shown to perform very well in assessing driving safety. The developed DSI model is intended to provide drivers and transportation authorities with timely and interpretable safety information, supporting informed decision-making to mitigate weather-related accident risks.

ABSTRACT. Digital environments increasingly influence how individuals access information, allocate attention, construct interpretations, and make decisions. Existing approaches frequently focus on content-level phenomena such as misinformation, persuasion, manipulation, or behavioral outcomes. Less attention has been given to the structural conditions under which judgment occurs and the recurring mechanisms through which those conditions may be altered.

This paper introduces the Interference Vector framework, a descriptive taxonomy for classifying structural forms of interference within digital cognitive environments. The framework is not intended as a governance or regulatory proposal. Its purpose is descriptive and classificatory. Historically, standards, measurement regimes, and governance frameworks have depended upon a shared vocabulary for identifying and describing the phenomena under consideration. Interference Vectors is intended as a contribution at that foundational descriptive layer.

Developed through a cross-domain synthesis of cognitive psychology, affective science, decision science, social and identity research, and systems-oriented analysis, the framework identifies twenty-two canonical Interference Vector Types organized across five analytic categories: attentional and temporal access, affective and motivational weighting, interpretive framing and meaning integration, identity and self-model regulation, and deliberative control and attribution.

The framework is intentionally non-normative and implementation-agnostic. Rather than evaluating truth, intent, or harm, it provides a structured vocabulary for identifying and comparing recurring interference mechanisms that shape the conditions supporting judgment and decision-making. Digital environments are used as the primary illustrative domain due to their scale, observability, and growing relevance to trust, resilience, and information integrity challenges.

By providing a common descriptive framework for analyzing structural interference, the taxonomy establishes a foundation for future empirical research on digital cognitive environments, socio-technical resilience, and the design and evaluation of systems that influence human judgment.

ABSTRACT. Air Traffic Management (ATM) increasingly relies on predictions of arrival times, delays, demand, and weather that are mapped to operational decisions through ranking, thresholding, and constrained allocation. Yet benchmarking often remains prediction-centric, using mean absolute error (MAE) or root mean squared error (RMSE), even when predictions drive safety-critical or cost-sensitive decisions. We propose a lightweight Decision-Driven Evaluation Protocol (DDEP) that pairs any predictor with a declared decision policy, constraints, and a cost structure, and reports decision cost and regret alongside predictive metrics. Using a reproducible European case study based on the EUROCONTROL Open Performance Data Initiative (OPDI) Flight List for June 2024, we map arrival-time predictions to greedy arrival-slot allocation with a separation constraint. We demonstrate a ranking inversion: the calibration minimizing MAE is not the calibration minimizing decision cost, for both a route-conditioned mean baseline and an embedding-based neural witness model. We explain the mechanism through a calibration trade-off curve and show that, under an asymmetric linear cost, the decision-optimal calibration is a quantile of an induced slack variable. Sensitivity analysis across cost ratios and separation settings confirms the robustness of the phenomenon across airports and model classes.