The purpose of this study is to examine whether analyst-level credit can be assigned quantitatively in a lightweight human-feedback decision-support pipeline. In intelligence and national security workflows, analysts often provide edits, comments, and evaluative feedback during the production of analytic products, yet these intermediate contributions are usually discarded, leaving no auditable record of how individual feedback shaped the final output. To address this problem, this study proposes a proof-of-concept Analyst-of-Record framework that combines synthetic analyst feedback, a linear ridge reward model, first-order influence functions, and additive Shapley aggregation to estimate both feedback-item and analyst-level contribution scores. The research design uses the Fact Extraction and VERification (FEVER) fact-verification dataset under controlled experimental settings. The pipeline retrieves evidence with Best Matching 25 (BM25), generates a grounded template-based response, derives three synthetic analyst feedback channels from FEVER annotations, trains a reward model on simple claim-answer and analyst-identity features, and aggregates per-feedback influence scores into an Analyst Contribution Index (ACI). The main experiments are conducted on a 500-claim subset across five random seeds, with additional ablation and bootstrap analyses used to assess sensitivity and stability. The findings show that the reward model achieves a mean validation R2 of 0.801 +/- 0.037, indicating that the synthetic feedback signals are learnable under the selected featureization. The analyst-level contribution scores remain stable across random seeds, with approximately half of the total influence magnitude attributed to the explanation-quality channel and the remainder split across the other two channels. Ablation results further show that removing the explanation-quality channel collapses validation fit, while bootstrap resampling demonstrates tight concentration of absolute ACI magnitudes. Theoretically, this study extends attribution research beyond document-only grounding by showing how analyst feedback itself can be modeled as an object of contribution analysis. It also demonstrates that influence functions and Shapley-style aggregation can be adapted into a tractable framework for estimating interpretable analyst-level credit in a reproducible experimental setting. Practically, the proposed framework offers an initial foundation for more traceable and accountable decision-support workflows in which intermediate analyst contributions can be preserved rather than lost. The results also provide a feasible implementation path for future systems that incorporate stronger generators, richer evidence representations, and real analyst annotations.

The rapid advancement of Artificial Intelligence (AI) systems has raised critical questions about their social, ethical, and practical implications. Many of these technologies are made by the few to represent the many, resulting in a considerable amount of harmful biases. While extensive research exists on AI bias and ethics, there lacks a comprehensive framework for understanding how culture, the foundation of bias, manifests in AI systems in ways that could impede the development of the emerging Artificial General Intelligence (AGI). This paper proposes aTi (pron. Ah Tee), a framework for holistically studying and quantifying cultural emergence in AI systems. We base aTi on four key components: values, behavioral patterns, decision algorithms, and experiences, which serve as activation points where culture can emerge. Our framework categorizes artificial intelligent systems into first-order and second-order artificial culture categories, offering a nuanced approach to understanding the complexity of cultural emergence in AI systems for AGI. Through case studies, we demonstrate the practical application of our framework and explore its implications for future AI development. This work contributes to the growing discourse on responsible AI development by providing a systematic approach to understanding and addressing cultural considerations in AI systems.

Joint base station (BS) association and beam selection in multi-UAV aerial corridors constitutes a challenging radio resource management (RRM) problem. It is driven by high-dimensional action spaces, need for substantial overhead to acquire global channel state information (CSI), rapidly varying propagation channels, and stringent latency requirements. Conventional combinatorial optimization methods, while near-optimal, are computationally prohibitive for real-time operation in such dynamic environments. While learning-based approaches can mitigate computational complexity and CSI overhead, the need for extensive site-specific (SS) datasets for model training remains a key challenge. To address these challenges, we develop a Digital Twin (DT)-enabled two-stage optimization framework that couples physics-based beam gain modeling with DRL for scalable online decision-making. In the first stage, a channel twin (CT) is constructed using a high-fidelity ray-tracing solver with geo-spatial contexts, and network information to capture SS propagation characteristics, and dual annealing algorithm is employed to precompute optimal transmission beam directions. In the second stage, a Multi-Head Proximal Policy Optimization (MH-PPO) agent, equipped with a scalable multi-head actor-critic architecture, is trained on the DT-generated channel dataset to directly map complex channel and beam states to jointly execute UAV-BS-beam association decisions. The proposed PPO agent achieves a 44%-121% improvement over DQN and 249%-807% gain over traditional heuristic based optimization schemes in a dense UAV scenario, while reducing inference latency by several orders of magnitude. These results demonstrate that DT-driven training pipelines can deliver high-performance, low-latency RRM policies tailored to SS deployments suitable for real-time resource management in next-generation aerial corridor networks.

Digital twins (DTs) can reduce over-the-air validation cost in industrial wireless networks, but their utility depends on the fidelity of the underlying channel twin (CT). At W-band, site-specific ray tracing captures deterministic propagation geometry, yet its channel frequency responses (CFRs) do not reproduce the small-scale impairments and capture-to-capture variability observed in laboratory orthogonal frequency-division multiplexing (OFDM) measurements above 90 GHz. This paper proposes Statistics-Consistent Sim-to-Lab Adaptation (SC-SLA), a calibration framework that improves the fidelity of a 95 GHz Sionna ray-traced CT toward that of the testbed by aligning the mean power delay profile (PDP), the distribution of root-mean-square delay spread (τ_rms), and per-subcarrier statistics at 50 MHz sampling bandwidth. SC-SLA uses a generative adversarial network (GAN)-inspired, cycle-consistent architecture with ResNet generators and batch-level channel-statistics losses on the PDP, sub-band PDP, τ_rms moments and quantiles, and normalized mean-square error (NMSE) of the mean CFR-magnitude profile. The framework is non-adversarial and requires neither paired simulated/measured samples nor discriminators. On held-out 95 GHz data, SC-SLA reduces the τ_rms Kolmogorov-Smirnov (KS) statistic from 0.86 to 0.050 relative to the impairment-augmented ray-traced input, and by 44

Artificial intelligence (AI) is undergoing a fundamental shift from task-specific assistance to collaborative partnership with humans. While deep learning has enabled significant advances in perception and prediction, its limitations in reasoning, explainability, and alignment with human intent are increasingly apparent. As AI evolves from a tool into a collaborator, next-generation systems must move beyond statistical pattern recognition to incorporate robust reasoning, transparent decision-making, and adaptability in dynamic environments. Neuro-symbolic AI (by integrating neural learning with symbolic reasoning) emerges as a foundational paradigm for this transformation. By combining the strengths of data-driven models with structured, interpretable reasoning, it enables shared cognition, more explainable outcomes, and closer alignment with human goals. This paper presents a future in which AI systems function as embedded partners across consumer electronics and smart environments, actively participating in human-centered cognitive processes. It examines the architectural foundations, interaction models, and domain-specific applications that will shape this transition, while also identifying the technical, ethical, and societal challenges that must be addressed to achieve scalable, efficient and symbiotic human-AI collaboration.

As edge computing devices for Internet of Things (IoT) applications become more capable, the demand for data-driven deep neural network (DNN)-based algorithms grows. However, this leads to increased model vulnerability to adversarial attack. The visual IoT (VIoT), a subdomain of IoT, has a great need for methods to improve the robustness of algorithms to such attacks, along with person re-identification (Re-ID) through the 3-D skeleton-based gait analysis. For other domains, either adversarial training or synthetic data augmentation has been shown to improve model performance in benign and adversarial attack cases. However, a combined approach that leverages both for robust classification in IoT has yet to be studied for 3-D skeleton-based Re-ID. In addition, assessment of 3-D skeleton-based Re-ID model vulnerability to adversarial attacks along with methods to improve robustness against such attacks is a critical literature gap. This work seeks to address these gaps by proposing an adversarial attack-tolerant computational framework for 3-D skeleton-based Re-ID. Our approach, for the first time in the literature, leverages both synthetic data augmentation using a generative adversarial network (GAN) and adversarial training through transfer learning of 3-D skeleton data on a convolutional neural network (CNN) for person Re-ID. We evaluate the robustness of our method against seven popular adversarial attacks across four unique Re-ID 3-D skeleton datasets. We demonstrate how our method responds to varying amounts of real data and study the viability of our method for deployment on edge devices. Our findings show that the proposed approach improves test accuracy in both the benign and adversarial cases. Our method is deployable to small and inexpensive edge devices popular in IoT such as the Raspberry Pi microcomputer.

The necessity for ultra-low-power consumption devices that enable the Internet of Things (IoT) networks to be more energy efficient, leads to considering Ambient Backscatter Communication (AmBC). Such devices exploit ambient radio-frequency signals to communicate and harvest energy. Though, the physical layer key generation (PLKG) scheme that enables AmBC to generate a secret key for data encryption, is vulnerable to injection attacks. Under such a menace, an attacker injects predefined signals and contributes to the PLKG. Thereby, it possesses a part of the generated bits and uses brute force attack to compromise the entire key. Inspired by that, we carefully prove the vulnerability of the existing PLKG approach in AmBC to injection attack. Then, we propose the employment of a smart device that aims at detecting the attack and canceling the malicious signal. Furthermore, we propose a new strategy that enables legitimate devices to utilize the entire received ambient signal for PLKG while exploiting the malicious signal for energy harvesting. Hence, the attacker becomes the victim. Finally, we provide numerical simulations that accentuate the consequence of the menace and the effectiveness of our security method, even though only 80% of the injected signal is removed. Besides, we prove the efficacy of our energy harvesting strategy.

Recognizing inattentive students is more complex on online platforms such as Zoom. The widespread adoption of virtual learning has increased the demand for reliable and interpretable methods to detect learner engagement. Although several models have been proposed, many methods still struggle to accurately capture engagement, often misclassifying inattentive students as engaged. To address this limitation, we propose a framework guided by robust visual representations, temporal modeling, and explainability to enhance engagement recognition in virtual learning environments and support academic success. Our approach combines facial image features extracted using DINOv2 with head pose parameters (yaw, pitch, roll), fusing these features to achieve more holistic engagement modeling.DINOv2 provides highly transferable embeddings and inherent attention maps that highlight visual regions, making it particularly effective for identifying facial cues of engagement. Video frames are scored based on these facial regions, with the top- K frames aggregated into a compact representation that is temporally modeled to capture dynamic cues before being classified into binary engagement labels. To ensure interpretability, explainable AI (XAI) is applied in post-hoc analysis, validating that model predictions align with key facial cues. Experiments conducted on the DAiSEE and EmotiW2023-EngageNet datasets demonstrate that our approach achieves 88% accuracy on EmotiW2023 and $\mathbf{7 5 \%}$ on DAiSEE, consistently outperforming existing models while enhancing transparency and trust.

Agentic AI is crossing trust boundaries faster than current risk models can represent. Existing approaches provide one of two partial views. They either describe failure mechanisms without producing a transferable residual-risk estimate, or they produce a risk estimate while treating the internal failure path as a black box. We couple those two views by proposing CPSAINT, a seven-layer integrity decomposition over Physical state, Sensors, Data, Compute, Actuators, Environment, and Time, paired with FRIESA-K, a residual-risk functional that maps each failure path to a quantified risk instance. FRIESA-K grounds the resistance term K in a controlled absorbing Markov model so that control effectiveness is derived from state dynamics rather than assigned as an informal score. The result is a concise mechanism-to-magnitude pipeline for resilient agentic and embodied AI. We report governance observability through a separate additive penalty instead of inserting governance as a new variable in the resistance functional. We formalize structural composability linking valid failure paths to well-defined risk instances and show the framework on two contrasting scenarios: a hard real-time warehouse robot and a governance-instrumented financial-services agent. Across both cases, the same layer grammar, variable semantics, and dynamic-resistance construction remain intact. Thus, we obtain a compact kernel that supports cross-domain reasoning, explicit assumptions, and quantitatively grounded formalism of composable trust.

Healthcare data breaches cost hospitals and clinics an average of $7.42M USD per breach. Autonomous multi-agent LLM pipelines in clinical settings multiply that exposure structurally: a prompt-injection payload embedded in a laboratory report or referral note traverses the full delegation hierarchy before any human reviewer observes it. Standard mechanisms including flat Role-Based Access Control (RBAC) and Macaroon caveat chains provide no structural guarantee that a spawned sub-agent's permissions remain bounded by its parent's scope. The Minimum Necessary Standard must hold across mixedsensitivity EHR data and multi-hop chains simultaneously. That is, the gap is architectural, not behavioral. We present the Hybrid Scripted-Execution and Prompt-Grounded Injection (SEPI) design, a replay-faithful methodology that freezes authorizationrequest traces and replays them identically across baselines to isolate authorization semantics from LLM stochasticity. Applied to a five-agent post-surgical oncology coordination workflow over a synthetic Electronic Health Record scenario, and evaluated against 39 clinically grounded injection payloads across leafagent, mid-delegate, and parent-orchestrator compromise scenarios, NGAC-hypergraph time-scoped delegation reduces mean sensitivity-weighted blast radius $\left(B_{\mathrm{w}}\right)$ by 75.0% relative to flat RBAC and 65.2% relative to Macaroon caveat chains. A temporal ablation shows that removing $t_{\text{expire }}$ raises mean $B_{\mathrm{w}}$ by 75.0% relative to full NGAC. The reduction follows from structural properties of the delegation architecture: attenuation is provable by construction rather than contingent on model behavior. Across model conditions, Claude Sonnet 4.6 actively detected all 39 payloads while open-source models showed weaker or intermittent behavioral protection, reinforcing authorization as the model-agnostic containment control.

To ensure robust connectivity in environments affected by signal blockage, simultaneously transmitting and reflecting reconfigurable intelligent surfaces (STAR-RIS) are introduced to provide seamless coverage to both indoor and outdoor users. Non-orthogonal multiple access (NOMA) is incorporated to increase spectral efficiency and serve multiple users. Under this configuration, the objective of this paper is to analyze the physical layer security (PLS) performance of STAR-RIS-assisted NOMA systems in scenarios involving multiple eavesdroppers. Specifically, the STAR-RIS reflects and transmits signals to near and far users for both indoor and outdoor propagation environments, which exposes the communication to potential interception by external adversaries. To assess the system’s vulnerability, closed-form intercept probability expressions are derived for both near and far users under colluding and non-colluding eavesdropping scenarios. The analysis further examines the influence of critical design parameters, including the number of transmitting and reflecting elements, multi-eavesdropping conditions, power allocation strategies, and fading characteristics, on the overall security performance. 1

Efficient resource management in the centralized unit user plane (CU-UP) of 5G and beyond networks is essential to support the diverse and dynamic demands of next-generation wireless Internet of Things (IoT) applications. In this article, we present a novel autoscaling technique that leverages the capabilities of the xApp and rApp features of the near-real-time radio access network (RAN) intelligent controller (near-RT RIC) and non-real-time RAN intelligent controller (non-RT RIC) to achieve sub-millisecond latency and broadband capabilities necessary for ultra-reliable low-latency communications (URLLC) and enhanced mobile broadband (eMBB) slicing use cases in the RAN. The proposed RICStar leverages three quality-of-service (QoS) prediction models: latency prediction, throughput estimation, and provisioning latency models to execute optimal capacity adjustment decisions based on the output of a time-series traffic forecasting model. Our evaluation uses a publicly available dataset representative of eMBB and URLLC traffic, where the autoscaler adjusts the capacity units available for each slice. Our results indicate that RICStar achieves better QoS and cost performance compared to canonical autoscaling techniques. This improvement is achieved by closely tracking traffic demand trends and proactively autoscaling CU-UP resources. These findings underscore the benefits of predictive autoscaling in mitigating network degradation caused by under-provisioning while reducing costs for communication service providers (CSPs). As a result, CSPs can readily support emerging applications such as telemedicine, connected vehicles, and augmented reality/virtual reality (AR/VR) with greater capacity and reliability.

Autonomous Vehicles (AVs) rely extensively on GPS signals for navigation, exposing them to a wide range of GPS spoofing attacks, from simplistic signal manipulation to sophisticated, coordinated falsification. Existing detection and mitigation solutions, both conventional and AI-based, face several critical limitations: they struggle to adapt to novel or evolving GPS spoofing strategies, rely on shallow or handcrafted features that fail to capture the semantic complexity of signal distortions, and often lack scalability, as they are typically designed for isolated scenarios and cannot be readily extended to heterogeneous AV fleets. In this study, we introduce a novel framework that integrates multiple Lightweight Language Models (LightLMs), including BERT, RoBERTa, DistilBERT, and TinyBERT, with Reinforcement Learning (RL) algorithms such as Q-Learning, Deep Q-Network, Advantage Actor-Critic, and Proximal Policy Optimization, to improve detection and mitigation of GPS spoofing. The LightLMs are used to convert structured GPS-related features into enriched state embeddings, which serve as input to the RL agent. These embeddings provide semantically meaningful representations that help the agent recognize complex spoofing behaviors and apply mitigation strategies. To train and evaluate the proposed models, we build a Python-based simulation environment that emulates multiple spoofing scenarios and integrates LightLM-driven state inputs. Experimental results across the datasets used show that RL models enhanced with LightLM-generated state representations significantly outperform their baseline counterparts in detection accuracy, mitigation efficiency, and response time. The results also demonstrate the proposed approach’s scalability, generalizability, and operational reliability for secure AV navigation.

This paper addresses the challenges of signal interception and detection in Terahertz (THz) networks. We explore how preserving the Gaussianity of transmit signals enhances Low Probability of Intercept and Detection (LPID) features using a data-driven autoencoder in THz band communication systems. By analyzing the probability distribution function (PDF) of THz band transmit signals, we evaluate their impact on LPID performance. Given the high sensitivity of THz systems to Phase Noise (PN) and the In-phase and Quadrature (IQ) imbalance impairments, we illustrate the importance of maintaining Gaussianity under varying PN conditions to ensure reliable and secure LPID transmission. Through comprehensive simulations, we demonstrate that PN does not compromise the Gaussianity of autoencoder-based THz band transmit signals. These results highlight the robustness of Gaussianity preservation, offering critical insights for designing THz band autoencoder-based LPID networks resilient to eavesdropping, jamming, and unauthorized interception. Our proposed autoencoder framework not only enhances LPID but also mitigates PN and IQ imbalance impairments, paving the way for secure, efficient, and reliable communication in next-generation THz networks.

Background: With advancements in Generative Artificial Intelligence, various industries have made substantial efforts to integrate this technology to enhance the efficiency and effectiveness of existing processes or identify potential weaknesses. Context, however, remains a crucial factor in leveraging intelligence, especially in high-stakes sectors such as healthcare, where contextual understanding can lead to life-changing outcomes. Objective: This research aims to develop a practical medical multi-agent system framework capable of automating appointment scheduling and triage classification, thus improving operational efficiency in healthcare settings. Methods: We present MedScrubCrew, a multi-agent framework integrating established technologies: Gale-Shapley stable matching algorithm for optimal patient-provider allocation, knowledge graphs for semantic compatibility profiling, and specialized large language model-based agents. The framework is designed to emulate the collaborative decision making processes typical of medical teams. Results: Our evaluation demonstrates that combining these components within a cohesive multi-agent architecture substantially enhances operational efficiency, task completeness, and contextual relevance in healthcare scheduling workflows. Conclusions:MedScrubCrew provides a practical, implementable blueprint for healthcare automation, addressing significant inefficiencies in real-world appointment scheduling and patient triage scenarios.

In the ever-changing world of online security, the increase in Internet of Things (IoT) devices has brought about new challenges in protecting the vast network of connected devices. Additionally, in the context of the escalating threat landscape within the IoT ecosystems, characterized by the growing complexity and sophistication of cyber threats, there exists a critical need for robust traffic detection mechanisms. In this paper, we combine the complementary strengths of three generative AI models, namely Autoencoders (AEs), Variational Autoencoders (VAEs), and Generative Adversarial Networks (GANs), to capture diverse anomaly patterns generated from IoT networks. Applying the ensemble models to the recent CICIoT2023 dataset generated from 115 IoT devices with 33 attack types, we measure the detection accuracy for benign and malicious traffic via majority voting as well as the weighted sum approach. Optimizing the threshold for the window size of 64, we achieve detection accuracy (recall) of 96 percent and 97 percent for normal and malicious traffic, respectively, outperforming recent state-of-the-art studies carried out on the same dataset.

Graph Neural Networks (GNN) have shown performance on learning on structured data and found applications in different fields like biology, physics, transportation, e-commerce. However, the convergence dynamic of GNN models, particularly the ones for link prediction tasks, remains a challenge for better design. Existing approaches have shown that some simpler GNN architectures can be effective as complex ones when used for recommender system (recsys). This encourages the need to perceive the advantage of some components towards a model convergence. We built models for link prediction, specially recsys based on a single layer GNN and evaluate their convergence on real and generated graphs data. These recsys are based on a graph convolution layer with added scaling factor or activation function. Our theoretical and numerical results highlight the advantage of a scaling factor over an activation function. It means models which scale down features before the final output converge, compared to others. Besides, the ReLU activation function alone cannot compensate for the missing scaling factor. The takeaway and advantage of a scaling factor resides in its ability to change the direction of the hidden features vector.

In smart manufacturing systems, Predictive Maintenance (PdM) is essential for optimizing equipment performance and reducing unplanned downtime for automated systems and machinery. Managing the vast amounts of real-time operational data generated by these systems requires a secure, scalable data-driven approach for robust PdM. This paper introduces a secure PdM framework employing Hyperbolic Cosine Differential Privacy (HCDP) to enhance attribute privacy in big datadriven predictive maintenance. Specifically, the system handles large-scale data acquisition, pre-processing, data grouping, and attribute extraction, with key attributes protected by HCDP. Next, the PdM analysis is conducted using Transfer Learning Generalized Collapsing Long Nuclear Logarithmic Entropy Short Term Memory (TL-GCLNLESTM), a model that leverages transfer learning to adapt to diverse scenarios and identify patterns within massive data streams ensuring predictive security. Then the Local Interpretable Model-agnostic Schwefel Explanation (LIMSE) framework interprets predictive outcomes, enhancing transparency and decision-making. If a failure is predicted/anticipated, the maintenance schedule is recomended using Ramp Multi-peak Gaussian Fuzzy (RMG-Fuzzy), with alerts issued accordingly. Experimental results demonstrate that the proposed model achieves high accuracy showcasing the efficacy of the proposed approach for PdM in smart manufacturing maintenance.

In this paper, we design a physical (PHY)-layer digital twin model, called PHY-Layer twin by leveraging ray-tracing (RT) and deep learning (DL) for a reconfigurable intelligent surface (RIS) assisted cell-free wireless communication system. Specifically, a recurrent neural network (RNN) driven low-complexity resource allocation scheme is studied to optimize system resources by interfacing with the proposed PHY-Layer twin. The proposed PHY-Layer twin-assisted resource allocation scheme jointly learns the correlation between channel gains obtained from RT and the real-world environment, and it allocates optimal transmit power for access points (APs) of the cell-free wireless networks. Numerical results demonstrate that the proposed scheme outperforms the related state-of-the-art approaches while operating with lower computational overhead.