In this paper, we introduce a 2-D leaky-wave antenna design with suppressed open stopband operating both in the near-and far-field, offering enhanced performance compared to existing designs. Two configurations, a double strip (DS) and a single strip (SS) bull's-eye (BE) leaky-wave antenna structure, are analyzed and compared. The DS launcher demonstrates superior broadband performance, maintaining efficient beam collimation across a wide frequency range in the near-field and enabling continuous beam scanning in the far-field. In contrast, the SS configuration exhibits an open stopband at 19.2 GHz and no continuous radiation. Both launchers employ a coaxial TMpolarized feed mechanism to ensure optimal power transfer, and dispersion curves show identical crossover frequencies for both BE antenna implementations. However, the suppression of the open stopband combined with the wideband capability of the DS configuration, highlights the potential of this design for nextgeneration high-frequency communication and radar systems operating in the near-or far-field.

The proliferation of ransomware attacks across global systems has necessitated the development of advanced detection techniques capable of adapting to increasingly sophisticated threat behaviors. The Deep Ransomware Fingerprint Mapping (DRFM) framework introduced in this work utilizes a novel approach, integrating behavioral and binary fingerprinting within a multi-layered machine learning architecture to provide robust, automated ransomware detection. DRFM leverages convolutional and recurrent neural networks alongside ensemble techniques, enabling the system to capture intricate patterns associated with ransomware activity, including novel and obfuscated variants, through autonomous classification mechanisms. Experimental evaluations demonstrate DRFM's high detection accuracy across known and zero-day ransomware variants, with superior adaptability to evolving encryption and evasion tactics compared to traditional models. Additionally, the framework's consistent memory utilization and low-latency operation under heavy data loads demonstrate its practicality for real-time, largescale cybersecurity applications. DRFM's effectiveness in reducing false positive rates, coupled with efficient resource utilization, highlights its value as a scalable solution capable of enhancing proactive cybersecurity defenses. Through its comprehensive, autonomous design, DRFM establishes a significant advancement in ransomware detection, supporting the increased demand for reliable, automated systems in today's cybersecurity landscape.

Achilles tendon overuse injuries are common for long-distance runners. Ankle exos (exoskeletons and exosuits) are wearable devices that can reduce Achilles tendon loading and could potentially aid in the rehabilitation or prevention of these injuries by helping to mitigate and control tissue loading. However, most ankle exos are confined to controlled lab testing and are not practical to use in real-world running. Here, we present the design of an unpowered ankle exo aimed at reducing the load on the Achilles tendon during running while also overcoming key usability challenges for runners outside the lab. We fabricated a 500-gram ankle exo prototype that attaches to the outside of a running shoe. We then evaluated its reliability, acceptability, transparency during swing phase, and Achilles tendon offloading during treadmill and outdoor running tests. We found that the exo prototype reliably assisted 95-99% of running steps during indoor and outdoor tests, was deemed acceptable by more than 80% of runners in terms of comfort and feel and did not impede natural ankle dorsiflexion during leg swing for 86% of runners. The exo reduced peak Achilles tendon loads for most participants during running; however, reductions varied considerably, between near zero and 12%, depending on the participant, condition (speed and slope) and the precise tendon load metric used. This next-generation ankle exo concept could open new possibilities for longitudinal and real-world research on runners, or when transitioning into the return-to-sport phase after an Achilles tendon injury. 

The functionality of a statistical framework for unified multivariate vital signs data fusion is seen in its flexibility for switching between the two data spaces. This paper develops data recovery methods for data fusion built on composite similarity measures using three distinct variable specific weighting methods - mean, median, and Orthogonalized Gnanadesikan-Kettenring (OGK). Using spatial covariance parameters derived from empirical patient data, the method successfully exhibits the ability to accurately recover eight different vital signs from its fused counterpart. Performance evaluation using Root Mean Squared Recovery Error (RMSRE), Root Mean Absolute Recovery Error (RMARE), and Root Standardized Mean Absolute Fusion Error (RSMARE) demonstrated that all three weighting approaches achieved comparable and highly accurate results, with parameters converging within ±0.01 of 1.0. The recovered signals closely matched the original data patterns, preserving both long-term trends and short-term fluctuations. Notably, the method proved computationally efficient and robust across different weighting approaches, suggesting broad applicability in clinical settings. This approach offers a promising solution for dimensionality reduction in complex physiological datasets while maintaining clinical relevance, with potential applications in patient monitoring systems and physiological modelling.

In this paper, we extend our previously proposed Adaptive Continuous Adversarial Training (ACAT) method beyond the problem-space of SPAM filters to encompass the feature-space of Machine Learning (ML)-based Network Intrusion Detection Systems (NIDS). ACAT continuously improves model robustness by incorporating adversarial samples detected from the ML-NIDS input network traffic into the training process. Problem-space evasion attacks rely on inverse-feature mapping, where modifications to real-world objects result in targeted perturbations within the feature vector after feature extraction. Therefore, protecting ML systems against feature-space attacks inherently provides defense against both feature-space and problem-space evasion attacks. Our results demonstrate that ACAT significantly reduces adversarial sample detection time compared to traditional techniques. Moreover, the performance of the ML-based NIDS under attack improved dramatically, with accuracy increasing from 50.92% to over 99% after only three retraining sessions.

The release of the 3.5 GHz Citizens Broadband Radio Service (CBRS) band has unlocked significant opportunities for a diverse range of stakeholders, particularly enabling communities and municipalities to establish private wireless networks. Successful deployment of these networks and effective spectrum sharing among CBRS users require accurate and reliable measurements of propagation loss. Often, private stakeholders depend on expensive commercial software that uses analytical models, which may not accurately reflect specific network conditions, thus potentially leading to measurement inaccuracies. Additionally, CBRS network propagation is complicated by its high sensitivity to weather and foliage, which makes network planning, coverage, and interference assessment more challenging. To this end, this paper presents a framework and methodology for characterizing CBRS network propagation, incorporating an extensive measurement campaign within a live CBRS private network in Buffalo, NY. Our study focuses on experimentally measuring path loss, comparing it with theoretical models, and assessing the impacts of weather and foliage on network performance. The methodologies and insights detailed in this paper can guide researchers and engineers in improving network design and operational efficiency across various network environments.

Multi-antenna technology has been leveraged in the past decades to meet the continuous demand for increased system capacity. User densification and high data rate provision further necessitates the efficient and fair allocation of resources among the users to reap the full benefit of multi-antenna systems. Conventionally, the resource allocation has been optimized using channel state information of users, which poses a performance trade-off considering large overhead in dense system configuration. In contrast, the position information carries the potential for efficient resource allocation with minimal acquisition cost, and can be enhanced by utilizing the power of machine learning (ML) frameworks. This work focuses on applying ML model for fair resource allocation based on position coordinates of the users in the system. Specifically, we evaluate the performance of our previously proposed coordinates-based resource allocation scheme through ML [1] in an interference-limited scenario, along with a fairness metric applied for optimized resource allocation. Results show that coordinatesbased resource allocation through ML can be successfully used for both noise-limited and interference-limited systems, and performs consistently well under different propagation channel conditions. Surprisingly, the learning frameworks used for coordinates-based resource allocation require similar training time, despite the complexity arising due to interference consideration, and has a moderate storage footprint as an implementation cost in the system.

As the network continues to become more complex due to the increased number of devices and ubiquitous connectivity, the trend is shifting from a centralized implementation to decentralization. Similarly, strategies to secure networks are increasingly leaning towards decentralization for its potential to enhance security in future networks with the help of Machine Learning (ML) techniques. In this regard, Distributed Machine Learning (DML) techniques, such as Federated Learning (FL) and Split Learning (SL), are at the forefront of this shift, offering collaborative learning capabilities across network nodes while maintaining data privacy. However, ML requires vast amounts of dedicated computing, memory, bandwidth, and as a consequence, energy resources. Moreover, resource consumption ML techniques used for network security have mostly been overlooked, which presents a glaring challenge for future networks in terms of overall resource utilization. This research emphasizes the importance of understanding the resource consumption patterns of two important DML techniques, i.e., FL and SL, to analyze the consumption of critical resources when deployed for network security. Furthermore, this research draws important insights from a practical comparative analysis of FL and SL in terms of resource consumption patterns and discusses their scope for future network security, such as in 6G, and stirs further research in this area.

Molecular communication (MC) is envisioned to realize nanotheranostics as an emerging diagnostic tool to improve existing treatment modalities. Phase separation (PS) is a complex time-dependent process responsible for discrimination of two independent phases from a single homogeneous mixture. Recently, PS was revealed to be the fundamental mechanism behind formation and organization of the living cells. Inspired by PS mechanisms in nature, we establish a novel modulation scheme for MC which encodes the information in the dispersion of molecules diffusing in the environment. Hence, the diffusion distribution can be considered as carriers of molecules. To evaluate the performance of this communication scheme, a dualcarrier diffusion-division modulation (DDM) is adopted where each carrier can be effectively modeled by superposition of multiple independent phases. We derive the theoretical bit error rate (BER) of the proposed DDM scheme which is validated by particle-based simulation (PBS). Furthermore, performance of the multi-carrier DDM scheme is compared to the well-known on-off keying (OOK) modulation scheme (as the most relevant benchmark) and pulse-position modulation (PPM) scheme. Interestingly, the proposed DDM is a covert modulation scheme since any other receiver cannot decode the transmitted signal by just counting the number of received molecules unless having the shared key. Moreover, the DDM scheme requires a receiver that exploits the displacement distribution of the molecules inside the receiver to infer about the tranmitted bit. We strongly believe that this concept will introduce novel types of communication schemes more compatible with biological microenvironments. Also, this work establishes the foundation for more complex orthogonal multi-carrier DDM schemes which can potentially unlock novel cellular sensing mechanisms in biology.

The capacity for sustained contextual coherence in language model outputs remains a critical factor in achieving natural, relevant, and semantically rich responses across extended interactions. Traditional context-handling methods, often limited to token-level attention and basic memory retention mechanisms, fall short when confronted with the intricate, hierarchical nature of real-world discourse. Multi-Stage Latent Contextual Synthesis presents a novel solution, employing a structured, multi-tiered synthesis approach to refine context integration within an LLM framework. Through a sequential layering process, latent knowledge is synthesized at progressively deeper levels, yielding a more nuanced understanding of semantic dependencies, thematic continuity, and response relevance across multi-turn conversations. The experimental results reveal substantial improvements in contextual retention, thematic alignment, and token stability, underscoring the method's efficacy in enhancing complex context management without adding prohibitive computational demands. Comprehensive evaluations demonstrated that Multi-Stage Latent Contextual Synthesis not only enhances memory efficiency and reduces thematic error rates but also promotes embedding stability, thereby supporting a more structured and responsive approach to contextual synthesis in high-stakes applications, including conversational AI and automated support systems. These findings illustrate the potential of hierarchical, latentcontextual strategies to redefine contextual processing standards within LLMs.

This letter introduces a novel approach for online bad data detection in distribution system state estimation (DSSE) by integrating compressive sensing (CS) with a modified largest normalized residual (LNR)-based detector. To the best of the authors' knowledge, this is the first work to develop a bad data detection method specifically for CS-based DSSE in unobservable distribution networks. The paper derives a closed-form solution for the compressed sensing problem, which is then used to quantify the error statistics in CS-based DSSE estimates. These statistics enable the design of a modified LNR-based detector using eigen decomposition, significantly improving anomaly detection. Extensive simulations on IEEE 37-bus and 123-bus unbalanced distribution systems demonstrate that the proposed method consistently outperforms the conventional LNR approach, achieving superior detection rates even with a limited number of measurements. This robust approach effectively detects data anomalies from both random errors and cyber-attacks, making it highly suitable for practical DSSE applications.

This paper presents FORTUNE, a hardware-agnostic fault tolerance technique for DNNs that leverages quantization to enhance reliability without significant performance overhead. Unlike conventional methods like Triple Modular Redundancy (TMR), which are computationally expensive, the proposed approach uses memory savings from quantization to protect the critical Most Significant Bit, improving fault tolerance in Deep Neural Networks (DNNs). Memory utilization has been reduced by 37.5% across all networks, with vulnerability in AlexNet reduced by 56% compared to the 8-bit version and 84% compared to the unprotected 3-bit version. These improvements come with only a minor increase in execution time of less than 3%. Using AlexNet as an example demonstrates how our approach effectively enhances memory utilization and resilience while causing only a minimal increase in execution time.

This paper presents the design, fabrication, and measurement of a compact band-stop filter unit operating in the subTHz F-band (90-140 GHz). The unit utilizes a λ/4 open-end folded stub resonator structure integrated within a coplanar waveguide (CPW). Different variants of coupling between the stub and CPW were analyzed using frequency domain simulations. The unit is implemented on an optically transparent quartz glass substrate, compatible with advanced optical microscopy techniques. This feature enables future applications in structural biology thanks to straightforward integration with a microfluidic channel. Another possible application is in 5G and upcoming 6G systems because the easily tunable filter unit design emphasizes electrically small size and manufacturing simplicity. The unit could be cascaded into a higher-order filter to enhance its performance.

Generating executable code from natural language instructions using Large Language Models (LLMs) poses challenges such as semantic ambiguity and understanding task-specific contexts. To address these issues, we propose a system called _DemoCraft_, which enhances code generation by leveraging in-context learning and demonstration selection, combined with latent concept learning. Latent concept learning introduces additional concept tokens, which are trainable embeddings that capture task-specific knowledge. We then test our system on two major datasets: _MBPP_ and _Humaneval_. Our experimental results demonstrate that the proposed system achieves an approximate _2x_ increase in the _pass@k_ metric compared to baseline models. Furthermore, we introduce two novel evaluation metrics: _correctness@k_ and _similarity@k_. Our empirical studies indicate that our system attains nearly a _3x_ improvement in these metrics as well.

Artificial Intelligence, specifically, Boost-base methods, has successfully revolutionized the area of stock price forecasting in recent years. This article integrates the idea of the hybrid stacking ensemble learning to forecast the trend of stocks. In the utilized Google stock price dataset, the market features "Open," "High", "Low," and "Volume" are utilized as input variables for a proposed hybrid stacking model in line with various Artificial Intelligence (AI) techniques. Unlike traditional stock price forecasting methods, this hybrid stacking approach incorporates multiple AI models, including XGBoost, CatBoost, and Light Gradient-Boosting Machine (LightGBM), with Lasso regression serving as the meta-learner. By employing the intrinsic pattern hidden in the real-world stock market and the inherent stacking mode of the proposed methodology, the hybrid algorithm in this paper endeavors to enhance the prediction capability and capture the real-world operation mode in financial market. By leveraging distributed stock price data inputs and the stacking methodology, the model aims to enhance forecasting accuracy and better reflect actual stock market conditions. The multistate algorithm outperformed on the widely-employed Google datasets with several typical prediction evaluation criteria, including some classical format. The benchmark experiments on various baseline models demonstrate the effectiveness and the advantages of the developed multi-stage algorithm and enjoy the significant improvement on the Google standard market competition.

This paper discusses a method to optimize the Transformer model using the Arctic Puffin Optimization Algorithm (APO) for predicting and classifying users' emotional states. The experiment is divided into a main experiment and a comparison experiment. In the main experiment, the APO-optimized Transformer shows significant performance improvement, with the fitness curve stabilizing after two iterations. The model achieved 98.79% accuracy on the training set and 98.10% on the test set, with the accuracy and loss curves further validating its effectiveness. In the comparison experiment, we compared the proposed method with three basic machine learning models, with the random forest model performing best at 94.7% accuracy. However, the APO-optimized Transformer surpassed the random forest, achieving a 3.4% improvement with 98.10% accuracy. These results demonstrate the potential of the APO algorithm in deep learning and its effectiveness for social media sentiment analysis.

QUIC, a general-purpose transport protocol, has recently been given consideration as a candidate solution for the Internet of Things (IoT). Researchers have proposed variants of application-layer protocols such as MQTT, CoAP, and AMQP all using QUIC transport. In this paper, we aim to show that since QUIC is so feature-rich on its own, many of the upper-layer complexities of traditional protocols can be greatly reduced or even eliminated. To exemplify this, we have designed what we call a 'QUIC thin-app' for IoT relying almost entirely on QUIC's specification. We achieved comparable functionality to MQTT, doing away with the need for much of its control messaging. Furthermore, our experimentation with our solution against MQTTv5 shows significant reductions in signaling overhead while maintaining comparable CPU and memory utilization.

Innovation and creativity are the most essential requirements for making education effective. Several innovative and creative tools and techniques have been developed and used for making education effective. Progressively, AI chatbots are another means by which educational institutions are able to provide an innovative, creative and interactive learning environment. AI chatbots are not just providing an innovative learning environment to students, but they are also providing help in almost all educational activities. In education, AI chatbots provide several benefits to individuals and educational institutions, however, there are several barriers in their use as well, and it is important to deliberate all these barriers when using these AI chatbots in education. Therefore, this paper will investigate several benefits and barriers of AI chatbots in education. Firstly, it will discuss the evolution of chatbots, and the use of some modern AI chatbots in education. Subsequently, it will investigate several benefits of AI chatbots in education. Finally, it will investigate several barriers of AI chatbots in education. This investigation will provide a detailed insight into both positive and negative aspects of AI chatbots alongside its use in education for better understanding.