Artificial intelligence (AI) and machine learning (ML) are widely adopted in sixth generation (6G) mobile networks. However, the deployment of AI in communication networks will require huge amounts of resources, such as computing, memory, bandwidth, and, as a result, energy. Certain use cases that are associated with resource-constrained devices, for instance, the internet of things (IoT), necessitate designing resource-aware and adaptable AI/ML techniques. In this article, a decentralized energy-aware federated learning (FL) model is proposed for IoT devices that allows the deployment of AI-based cybersecurity operations in 6G. We employ an ordered dropout (OD) mechanism to construct nested submodels from a larger neural network (NN), enabling dynamic adaptation to the energy availability of the system and reducing the overall energy footprint. The experimental evaluations show that the proposed energy-aware model extends the operational lifetime of the deployment framework from 82% to 135% for different datasets, while reducing inference time per sample by up to 50% for the smallest submodel.

Dynamic studies of the components involved in the interconnected power and hydrogen (H2) sectors are crucial for ensuring the security and stability of both sectors. The connection between electrolyzer plants and electric-driven compressor stations (EDCSs) is a key interaction between these sectors. This work, for the first time, explores the dynamics of a coupled electrolyzer-EDCS system. The developed model analyzes the dynamics of the integrated system when the compressor station is faced with power grid disturbances. Simulation outcomes reveal that the coupled electrolyzer-EDCS system works perfectly in normal conditions. However, the system is unable to appropriately respond to unexpected changes when the EDCS faces a disturbance from the power sector side. The identified issue is due to the lack of a proper control link between the electrolyzer and EDCS. To enable the electrolyzer for regulating its H2 production when the EDCS is disturbed, there is a need for designing of an effective control mechanism. Our verified developed system can be used for this purpose.

The feedback control of a quadrotor uncrewed aerial vehicle's (UAV) rigid-body is tightly coupled with the underlying actuation system, including the actuator dynamics and aerodynamics, the actuator placement's geometric configuration, and the control allocation algorithm. Taking these couplings into account, this paper analyzes the closed-loop stability of a quadrotor UAV governed by a cascaded control architecture and a static mixer-based control allocator-an approach widely used in open-source autopilot platforms such as PX4 and ArduPilot.

Safety compliance in high-risk industrial domains such as construction and mining remains persistent challenge due to the complex interaction of workers, heavy machinery, and complex site conditions. Conventional computer vision methods can detects objects or classify activities but often lack the semantic reasoning needed for comprehensive safety assessment. This paper presents SafeVision a unified vision-language framework for real-time, context-aware safety assessment. The SafeVision integrates a frozen CLIP-ViT encoder with two complementary reasoning modules: the Scene-Aware Reasoner (SAR) for global environmental context understanding and Region-Focused Neural Reasoner for region-level safety compliance verification. A task-conditioned visual question answering component further enables natural language queries regarding hazards, personal protective equipment (PPE) usage, and operational conditions. To support deployment on resource-constrained platforms, SafeVision parameter-efficient Low-Rank Adaptation (LoRA), ensuring scalability without loss of interpretability. Experiments on a custom multimodal dataset comprising 10,000 annotated images demonstrate strong performance across multiple tasks, with SAR reaching 87.4% accuracy in scene classification and the VQA component achieving BLEU-4 of 83.2%, F1 of 84.6%, Exact Match of 80.3%, and METEOR of 82.5%. Qualitative analysis highlights the effective alignment of global scene understanding and localized reasoning, underscoring the potential of SafeVision for proactive safety monitoring in industrial environments.

Diabetes is one of the leading causes of death and chronic illness in the United States. Invasive glucose monitoring reduces compliance, while current noninvasive methods lack sufficient accuracy for home use. In this study, we developed and preliminarily validated a low-cost, noninvasive glucose monitor using near-infrared (NIR) spectroscopy and machine learning. The prototype sequentially measured light reflectance from the index finger at 850 nm, 1450nm, and 1550 nm; reflected signals were then baseline-corrected with water-referenced differences to attenuate water absorption effects. From 130 paired measurements with a capillary reference, we trained five machine learning models, with support vector regression selected for best performance on 30 samples (R 2 = 0.95; RMSE = 5.39 mg/dL; MAPE = 3.93%). All predictions lay within Zone A of the Clarke Error Grid, and a two one-side t test (TOST) indicated the mean difference lay within ±15 mg/dL (P <0.001). This proof-ofconcept study demonstrates the feasibility of multispectral NIR with differencebased water correction and SVR for accurate noninvasive glucose predictions, and the potential for future multi-participant, range-expanded studies to test clinical utility.

Although the energy footprint of cloud infrastructures is becoming a critical concern, mainstream orchestration platforms such as Kubernetes (K8s) continue to prioritize performance over power efficiency. By default, K8s employs static CPU pinning for latency-sensitive services to ensure predictable performance and prevent resource contention. However, during service idle periods, residual background activities-such as runtime threads and maintenance tasks-occupy the assigned cores, preventing them from entering deep idle states (e.g., C6) and leading to unnecessary power consumption. This paper introduces EC6, a lightweight, idle-aware scheduling policy for K8s that consolidates idle services onto a reserved single CPU core, allowing other cores to transition into deep C-states during service inactivity. Upon resumption of service activity, EC6 transparently restores services to their original CPU cores with minimal latency, ensuring resource allocation constraints (i.e., , number of CPU cores) are maintained. Experimental evaluations show that EC6 increases per-core C6 residency by up to 30% and reduces idle power consumption by up to 13.6% compared to K8s' default container scheduling policy, without compromising performance. This demonstrates its effectiveness in improving energy proportionality in cloud-native environments while preserving service quality.

Edge platforms face tight latency, memory, and energy limits, often without accelerators [1], [3], [4]. Selecting accurate yet efficient models remains empirical. This letter compares lightweight vision pipelines on the Raspberry Pi 5-classical (HOG + logistic regression, decision trees) [6], [7] and compact CNNs (TinyCNN, MobileNetV1-α0.25) [9]-using a reproducible Colab→Pi workflow with post-training INT8 quantization [17], [18]. Accuracy, size, and batch-1 latency are measured on both workstation proxies and Pi 5 (2 GB/8 GB). TinyCNN surpasses classical baselines with sub-MB binaries; quantization yields 3-7× compression. Notably, FP32 kernels outperform INT8 at small scale due to runtime overheads and partial kernel coverage [24], [25]. Released artifacts enable exact reruns and clarify how architecture, quantization, and backend maturity jointly shape embedded inference performance.

The tradeoff between internet of things (IoT) network growth and user requirements significantly maximized which affected by the dynamically devolving malware attacks. The conventional security systems have utilized for scalable and intelligent data security, especially intrusion detection systems have used for IoT network by using the different network parameters. Several recent solutions also utilized for IoT network to address the problems, this work present the cryptography with transfer learning model for malware detection in IoT network. Here, the dynamic chaos encryption model is used for encryption purpose which generates the encrypted data with proper key generation process to ensure the data protection. The transfer learning model evolved from the fine-tuned rule-set includes convolutional neural network (CNN) and recurrent neural network (RNN) is used for network traffic patter analysis and detect the malicious types in the IoT network. The proposed work utilizes the Bot-IoT and TON_IoT datasets to validate the performance. The Chaos+RNN model achieved 99.624% accuracy on the Bot-IoT dataset, outperforming the previous best by 1.272%, with precision, recall, and specificity improvements of 1.301%, 1.308%, and 1.055%, respectively. On the TON_IoT dataset, it reached 99.472% accuracy, exceeding the top model by 1.365%, with respective gains of 1.447%, 1.449%, and 1.168%. The findings show that the system achieves superior malware detection performance, with increased precision and recall across various attack scenarios.

This paper investigates the use of short-term forecasts for microgrid operation under uncertainties of the renewable generation. Using hourly data from Ontario's K0K region in 2018, three representative months, January, July and December, were chosen to reflect seasonal conditions. We trained different univariate GRU models for load, photovoltaic and wind generation with the input window of 24 hours. We tested the predictions in a dual-branch MATLAB/Simulink setup, where one branch ran on actual profiles a nd t he o ther o n GRUbased forecasts, both controlled by the same BESS system. System performance was analyzed through grid export, battery discharge, and state of charge (SoC) and provided in the form of key performance indicators (KPIs). The GRU models provided reasonable statistical fits (R² of 0.78-0.96 for all sources and seasons). More importantly, stability-reflecting K PIs l ike g rid peak demand and average SoC were largely unaffected by forecastdriven operation, with only minor variations, while energy flowreflecting K PIs s uch a s g rid e xport a nd B ESS t hroughput were more impacted. This shows that GRU-based forecasts can be trusted for hour-ahead microgrid operation, since keeping the system stable matters more than fully correcting forecast errors.

This paper addresses a fundamental open problem in Large Language Models (LLMs): Are Chain-of-Thought (CoT) reasoning steps causally integrated into the problem-solving mechanism, or are they merely post-hoc rationalizations? I propose a formal mathematical framework using Structural Causal Models (SCMs) and Causal Mediation Analysis to rigorously define and quantify CoT faithfulness. I formalize CoT as a mediator (M) between the input prompt (X) and the final output (Y), explicitly accounting for biases introduced by latent confounders (U). My core contribution lies in three novel theorems that establish the conditions for the causal identifiability of CoT effects, provide bounds on bias when confounding is present, and demonstrate a faithfulness-robustness tradeoff. This work provides the first formal mathematical foundation for CoT faithfulness, offering a theoretical blueprint for future empirical validation in mechanistic interpretability.

Rural logistics continues to face persistent challenges, including high infrastructure costs, frequent road accidents, toll fees, and inefficient routing practices. Classical single-period inventory-routing models fail to adapt routing attributes (transportation costs, risks, routeaccess tolls, and distances) across time, and they rarely integrate delivery and pickup visits. Integrating pickup and delivery-where vehicles collect goods on return trips after making deliveries-is critical to reduce empty miles. This paper presents a novel mathematical model for the multi-period inventory pickup and delivery problem with multi-attribute alternative routes (MP-IPD-MAAR) in rural settings. The bi-objective formulation jointly minimizes total cost and accident risk, quantifying trade-offs between transportation expenditures and crash-related expenses using crash and transport data. A tailored Non-dominated Sorting Genetic Algorithm II (NSGA-II), enhanced with AI-driven fuzzy logic parameter tuning, solves medium-to large-scale instances and achieves 7-14% reductions in both costs and risks in our experiments. Based on the findings, among the three metaheuristic methods, Modified NSGA II, Standard NSGA II, and MOEA/D, the Modified NSGA II provided the best overall results.

In this paper, we revisit the well-known problem of Maximum Likelihood estimation of the positive definite Toeplitz covariance matrix for complex Gaussian random data. Specifically, we treat the problem of likelihood maximization as convex if the optimized likelihood exceeds the likelihood of the true covariance matrix, and we get the same solution, irrespective of the initial solution. Yet, to reveal this property, which is not limited by computational accuracy, the parameterization of the optimized Toeplitz matrix and the optimization algorithm should be selected appropriately.

The rise in wireless devices has made efficient spectrum management a critical challenge. Cognitive Radio Networks (CRNs) address this issue by allowing secondary users (SUs) to utilize vacant spectrum bands without causing interference to primary users (PUs). This study investigates the use of reinforcement learning (RL) techniques to enable smarter channel selection in CRNs. We begin with a tabular Q-Learning model for a single SU and check its performance in both static and dynamic environments, observing steady improvements in minimizing collisions. To handle larger and more complex state spaces, the project is extended to a Deep Q-Network (DQN) approach. Results of this experiment show that while DQNs converge faster and perform effectively for single user scenarios, their efficiency decreases in multi user settings, where collisions among secondary users still occur. These results show the challenges of applying RL methods to multi-agent CRNs and point to the need for more coordinated learning algorithms and strategies. Overall, the work highlights both the potential and current limitations of RL-based spectrum access in dynamic cognitive radio systems.

Quality control in manufacturing traditionally relies on error-prone manual inspections. This paper introduces an automated framework for industrial anomaly detection that requires only defect-free images for training. Our approach leverages convolutional autoencoders whose architecture and hyperparameters are optimized via Bayesian search to maximize a novel objective function. This function quantifies the separation between reconstruction errors of defect-free images and synthetically generated defects using diffusion models and classical augmentations-enabling robust semi-supervised learning without real defect data. The pipeline generates interpretable error maps for real-time operator feedback. Evaluated on the Metal Nut MVTec Anomaly Detection Dataset, our method achieves a state-ofthe-art AUC of 0.87 and IoU of 0.37, outperforming handcrafted baselines by 1% and 42%, respectively. The framework´s efficiency and minimal setup requirements make it ideal for deployment in industrial vision systems, significantly reducing manual intervention and enhancing reliability.

Objective: Vagus Nerve Stimulation (VNS) is an established technique for treatment of various neurological conditions. The major limitation of conventional VNS is no spatial and temporal selectivity: the whole nerve is stimulated in a tonic manner leading to adverse effects and non-optimal treatment. On-demand closed-loop stimulation linked to seizure events would allow enhancement of treatment outcomes, reduction in side effects and lower energy consumption by an implanted stimulator. The purpose of this study is to identify a biomarker of seizures on the vagus nerve to enable such ondemand stimulation. Methods: We performed an in vivo investigation in anaesthetised pigs to demonstrate the feasibility of detecting a biomarker of seizures in both branches of cervical vagus nerves using multi-electrode nerve cuffs. The data was analysed using the velocity-selective recording technique, which allowed differentiating between nerve fibres conducting at different velocities, and between afferent and efferent fibres. Results: We demonstrated a significantly increased activation of fibres conducting at 7-15 m/s (Aδ, Aγ) on the right vagus nerve during seizures in one animal. While more experiments are required, the findings of the study are the essential milestone for the development of a novel closed-loop neuromodulation system for treatment of epilepsy.

This work presents a vectorial extension of the TR-MUSIC algorithm tailored for microwave breast imaging. By leveraging the full vectorial formulation of electromagnetic fields and incorporating the antenna radiation characteristics via a source-scattering-matrix formalism, a generalized imaging framework is developed that directly links vector-networkanalyzer measurements to the electric field distribution. The proposed methodology is applied to both simulated and experimental data from anthropomorphic breast phantoms. Results confirm the theoretical consistency of the vectorial approach and highlight its capacity to capture complex antenna-field interactions with the imaging scene. While the scalar approximation continues to deliver stable and reliable reconstructions under typical conditions, the full-antenna model may provide additional flexibility and accuracy in scenarios where the assumptions of scalar-field representation or dipole-antenna model are not fully met. The scalar model is thus interpreted as a special case of the more general vectorial formulation. The preliminary test results included in this article serve as proof-of-concept for the incorporation of vector-based electromagnetic field modeling into future clinical microwave imaging systems for non-invasive breast cancer detection.

Drones (also referred to as unmanned aerial vehicles) have attracted significant attention within smart city ecosystems due to their ability to detect crises promptly and provide real-time support for disaster management. However, the increasing risks of cyber and physical attacks, along with potential private data leaks, present substantial challenges to their deployment in smart city environments. In this paper, we propose a secure communication protocol that leverages a novel group authentication scheme (GAS), tailored for disaster management scenarios in smart cities. The proposed scheme enables both group key agreement among drones and individual session key establishment between each drone and the control center (CC). Hence, our protocol preserves data confidentiality, message integrity, and data privacy, and verifies the authenticity of communicating parties. We employ physically unclonable functions (PUFs) and reverse fuzzy extractors to withstand critical threats, including machine learning (ML)-based modeling attacks and physical attacks. Our protocol also uses lightweight cryptographic primitives, such as hash-based message authentication codes (HMAC) and exclusive-OR (XOR) operations to ensure efficiency. Comprehensive security analysis using the formal proof and the Scyther tool demonstrates that our scheme resists corresponding attacks and protects data privacy. Ultimately, performance analysis results confirm that the proposed protocol is efficient and feasible.

The AC optimal power flow (ACOPF) problem is fundamental for ensuring the economic and secure operation of power systems. In recent years, artificial neural network (ANN)-based methods have emerged as promising approaches for learning the mapping from load inputs to optimal generation solutions. However, these methods often function as "black boxes," offering limited transparency into their intermediate decision-making processes and making it difficult to interpret the underlying reasoning of the models. To address these challenges, we propose a direction field learning method (DFLM) for the ACOPF solution space, based on score matching and diffusion models. The direction field provides explicit information about the optimization direction, enabling intuitive and visual interpretation of the direction field learned by the ANN. This, in turn, enhances the transparency and credibility of the proposed method. Case studies on the IEEE 24-bus and 118-bus systems demonstrate the feasibility of the approach, showing that DFLM achieves performance comparable to state-of-the-art methods while offering significantly improved transparency.