Tactical air-ground wireless sensor networks (TAG-WSNs) are mission-critical wireless sensor networks (WSNs) that employ airborne sensor nodes (ASNs) to capture aerial sensory data during military operations, thereby overcoming the sensing coverage limitations of the ground network. However, intelligent jamming attacks on the network's links, coupled with the highly dynamic network topology, disrupt data communication and pose challenges for reliable routing. In this paper, we introduce a cross-layer (MAC-PHY) jamming framework that models the hostile characteristics of TAG-WSNs. Secondly, we propose a scalable federated deep reinforcement learning (FDRL)-enabled routing solution called FedRoute, which enables agents to build a shared routing model. To support jamming-resilient collaborative model training, we use multiple spatially distributed mobile robot nodes (MRNs) as parameter servers. In FedRoute, local DRL models are meta-trained with the routing agents' exploration data before federated averaging, resulting in meta-optimized regional routing models. Moreover, FedRoute empowers routing agents to discover quick and reliable routes in the presence of jamming attacks on acknowledgment (ACK), negative acknowledgment (NACK), and data packets. The proposed scheme outperforms benchmark algorithms in terms of expected transmission count, packet delivery ratio, end-to-end delay, and energy efficiency.

Machine learning (ML) has achieved remarkable advancements, with applications spanning various industries. However, traditional ML methods often require large amounts of computational resources, labeled data, and expensive hardware, making them less accessible for smaller organizations, researchers, and low-resource environments. Frugal machine learning (FML) refers to a set of techniques that aim to make ML models more efficient, cost-effective, and accessible while maintaining or even improving their performance. This paper presents a comprehensive survey on the emerging field of frugal machine learning, highlighting key approaches, challenges, and future research directions. We classify frugal techniques into categories such as model efficiency, data efficiency, and hardware efficiency, and provide an overview of notable methods and their applications. We also discuss open problems and potential avenues for advancing FML.

With rapid digitization and digitalization, drawing a fine line between the digital and the physical world has become nearly impossible. It has become essential more than ever to integrate all spheres of life into a single Digital Thread to address pressing challenges of modern society – accessible and inclusive healthcare in terms of equality, and equity. The techno-social advancements and mutual acceptance have enabled the infusion of digital models to simulate social settings with minimum resource utilization to make effective decisions. However, there is a significant gap in feeding back the models with appropriate changes in real-time. In other words, an active behavioral modeling of modern society is lacking, that influences community healthcare as a “whole”. By creating virtual replicas of physical systems, digital twins can enable real-time monitoring, simulation, and optimization of urban dynamics. This paper explores the potential of digital twins to promote inclusive healthcare for evolving smart cities. We argue that digital twins can be used to:Identify and address disparities in access to healthcare servicesFacilitate community participationSimulate the impact of urban policies and interventions on different groups of peopleAid policy making bodies for better access to healthcareThis paper proposes several proposals for stitching the actual and virtual societies using digital twins. Several discussed concepts within this framework envision an active, integrated, and synchronized community aware of data privacy and security. The proposal also provides high- level step-wise transitions that will enable this transformation.

Histopathologists are experiencing a digital revolution in their field thanks to the digitization of Whole Slide Images (WSIs), which are microscope slides of tissue that can measure gigapixels in size. With so much high resolution data at their disposal, computer vision techniques can now be used to automate laboratory processes, create visual standards, and increase analysis throughput, all of which reduce the workload of pathologists [1]. The ”gold” standard in neuropathology, particularly for Alzheimer’s Disease—is pathological diagnosis made by looking at White Matter Inclusions (WSIs) in brain tissue. Semi-quantitative scoring in accordance with the standards established by the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) is necessary in order to arrive at a pathology diagnosis. As to these criteria, the diagnosis of Alzheimer’s disease relies heavily on the density of neuritic (cored) Amyloid-β plaques. In order to support in the classification of Alzheimer’s disease, recent studies by Tang et al. [2] and Wong et al. [3] examine the use of Deep Learning (DL) algorithms to WSIs of brain regions that have been histochemically labelled. According to the results from [2], there is a link between semi-quantitative scoring performed by a skilled neuropathologist and scoring generated by a DL, coupled with strong precision and recall metrics. By using the methods of [2] to data from an alternative brain bank, Vizcarra et al. [4] validate it and discover a similar connection between DL and semi-quantitative scoring. The methodology’s successful validation shows potential for incorporating DL into the neuropathology pipeline. While this can shorten the time neuropathologists need to score WSIs, two studies [2], [3] point out a specific problem—the predominant Aβ morphologies are quite rare, with diffuse plaques predominating. Classified as cored plaques or Cerebral Amyloid Angiopathy (CAA), only 1-2% of candidate samples found using conventional computer vision algorithms meet these criteria. Consequently, the significantly imbalanced datasets that are employed may have a negative impact on the DL’s performance. Additionally, trying to increase minority class instances by extra annotation by highly qualified neuropathologists would be prohibitively costly and time-consuming. Therefore, we aim to address the problem of unbalanced datasets for Aβ morphology in brain tissue immune histochemically stained WSIs. As part of our approach, we want to investigate how minority examples of Aβ morphologies, like cored plaques and CAA, may be synthesized using generative modelling, namely Generative Adversarial Networks (GANs), in order to balance the dataset. By doing this, we might also shorten the amount of time needed to classify high-quality datasets. To the best of our knowledge, this is the initial attempt at generatively modeling for the distribution of Aβ morphologies in images. Our objective is to respond to the following questions:Is it possible to develop a GAN architecture that can synthesize unique, high-quality instances of Aβ morphologies that are identical to real samples?Does the GAN prevent the problem of memorization and is it able to sufficiently address the modes of the real data distribution?Are there ever instances of low quality? If so, is it possible to eliminate this through a selection process?Can we use datasets balanced by GAN-oversampling minority data to increase the reliability of downstream classifiers for Alzheimer’s disease?The paper is as follows in the next section we will see the background related to our study. In Section 3, the related works are presented. In Section 4, the materials and methods used are discussed. In Section 5, the experimental analysis along with results are presented and we conclude the paper in Section 6 with some conclusions and future works.

In this article, we propose a digital agent (DA)assisted network management framework for future sixth generation (6G) networks considering users' quality of experience (QoE). Particularly, a novel QoE metric is defined by incorporating the impact of user behavior dynamics and environment complexity on quality of service (QoS). A two-level DA architecture is developed to assist the QoE-driven network orchestration and slicing, respectively. To further improve the performance of proposed framework, three potential solutions are presented from the perspectives of DA data collection, network scheduling algorithm selection, and DA deployment. A case study demonstrates that the proposed framework can effectively improve users' QoE compared with benchmark schemes.

Emerging 5G-Advanced and 6G wireless networks are anticipated to support a wide array of services, including further enhanced mobile broadband (FeMBB) and extreme ultra-reliable low-latency communications (eURLLC), to meet diverse communication needs. The radio access network (RAN) slicing is a pivotal technology for enabling the delivery of these services on shared infrastructure, playing a particularly important role in 6G, where FeMBB and eURLLC services have different blocklengths. To meet varying quality of service (QoS) demands in next-generation networks, innovative multiple access techniques are required to improve interference management and optimize spectrum efficiency. Rate-splitting multiple access (RSMA) is an effective approach for achieving these objectives. This paper investigates the problem of Energy-efficient joint Resource block (RB) allocation and Power control (ERP) for the coexistence of FeMBB and eURLLC services in RSMA-based green communication networks. In this ERP problem, each FeMBB user is guaranteed a minimum data rate, while each eURLLC user must satisfy latency and reliability constraints. To address the ERP problem, we introduce a sub-optimal algorithm (SO-ERP) based on convex optimization. However, the SO-ERP algorithm has high computational complexity and requires approximations to convexify the original ERP problem, potentially moving the solution away from the optimum. To overcome these limitations, we propose a hybrid deep reinforcement learning (HDRL-ERP) algorithm that employs a dueling double deep Qnetwork for RB allocation and a deep deterministic policy gradient for power control. Simulation results are presented to illustrate the performance of the HDRL-ERP algorithm.

Recent droughts increase the risk of wildfires in the Mediterranean region, which in turn produce greenhouse gas (GHG) emissions from biomass burning, further exacerbating climate change. Accurate mapping of forest fires is urgently required to avoid underestimating GHG emissions in national GHG inventories and for fire management and prevention. In this study, we test a cumulative sum change detection algorithm on both Sentinel-1 and Sentinel-2 data in two test areas in the Sardinia region, Italy. The objective is to improve the mapping of two small-scale fires (less than 50 ha) undetected by the European Forest Fire Information System (EFFIS). Through the integration of multi-sensor information, the resulting maps are able to match more than 50% of the fieldmeasured burnt area. The maps also provide the exact timing of the fire events, matching the records from the Forest Service. Given its versatility and mapping potential, this methodology shows promise for application in other Mediterranean regions.

This paper presents a novel end-to-end (E2E) learning architecture for massive MIMO communication systems using complex-valued neural networks (CVNNs). Our approach leverages CVNNs to process complex signals directly, eliminating the need to split real and imaginary components, thereby preserving the natural structure of wireless signals. The proposed architecture integrates both the encoding and decoding stages, optimized for flat-fading Rayleigh channel conditions, focusing on improving transmission efficiency. A key contribution is the extension of the approach to multiuser MIMO scenarios, where the system is designed to orthogonalize data streams for several user equipment, improving spectral efficiency with federated learning. We show that it is possible to effectively transmit a number of data streams that exceed the channel matrix rank. Additionally, a power control mechanism based on regularization is introduced to ensure stable transmission power. The effectiveness of the proposed approach is rigorously validated through simulations across a range of scenarios, demonstrating significant improvements in the mutual information. The results are compared with theoretical limits and classical approaches, highlighting the potential of CVNN-based architectures for advancing future wireless communication systems in both single and multiuser contexts.

The Web of Things (WoT) attempts to overcome the Internet of Things (IoT) fragmentation and interoperability limitations by using standard web protocols. While WoT brings a huge potential for developing new and custom distributed applications, the number of system’s components and system complexity keeps growing and is ever more difficult to manage, along with an increase in resource consumption, which is a critical issue in constrained environments as the IoT. In order to tackle this issue, a framework named Self-Orchestrated Web of Things (SOrWoT) is proposed that integrates Hierarchical Finite State Machine (HFSM) constructs into WoT, allowing for the modeling of behavior in applications, breaking down complex systems, and further promoting changes in the architecting of components. A use case for applying SOrWoT’s constructs to a smart city energy management system is discussed, to showcase the framework’s practical application. Our proposed framework has the potential to improve resource efficiency, cut operational costs, and enhance overall performance by decomposing complex processes into smaller, manageable components. This modular and hierarchical strategy also has the potential to improve the system’s adaptability, resilience, and scalability.

Moving away from plain-text DNS communications, users now have the option of using encrypted DNS protocols for domain name resolutions. DNS-over-QUIC (DoQ) employs QUIC-the latest transport protocol-for encrypted communications between users and their recursive DNS servers. QUIC is also poised to become the foundation of our daily web browsing experience by replacing TCP with HTTPP/3, the latest version of the HTTP protocol. Traditional TCP-based web browsing is vulnerable to website fingerprinting (WFP) attacks that can identify the websites a user visits. The emergence of QUIC-based DNS and HTTP protocols raises an important question: are regular users better protected from WFP attacks when using these new protocols? To investigate this, we first collect and publicly release the first benchmark dataset of network traffic corresponding to real visits to QUIC-enabled websites while using DoQ for domain resolution. This dataset will help advance the research on WFP attacks and defenses. Second, we implement and evaluate the first WFP attack targeting the combined use of DoQ and HTTP/3 protocols by users by developing two transformer models tailored for WFP attacks. Finally, we conduct comprehensive experiments, which reveal that these models are effective in identifying uservisited websites, emphasizing the need for defensive measures. SUBMITTED TO IEEE COMMUNICATIONS MAGAZINE, CYBERSECURITY (IN INCUBATION), 03/12/2024

Recursive Model Refinement introduces a novel mechanism for iteratively enhancing latent knowledge representation and task-specific adaptability within transformer-based architectures. The framework operates through a structured feedback-driven methodology, enabling the recalibration of internal embeddings to better align with evolving contextual demands. Experiments reveal significant reductions in error rates across complex linguistic tasks, including translation and summarization, through improved contextual coherence and token-level alignment. A mathematical formalization underpins the process, ensuring scalability and computational efficiency without compromising foundational pre-trained capacities. Comparative analyses highlight consistent performance gains relative to baseline methods, with observable benefits in semantic drift mitigation and vocabulary utilization rates across domainspecific benchmarks. The integration of recursive mechanisms into pre-trained models effectively balances generalization and specificity, a critical requirement for diverse natural language applications. Results demonstrate improved stability in latent space organization, reflected in clustering metrics and reduced representational discrepancies. Error analysis provides further evidence of the framework's capacity to systematically address inherent limitations in static representations. Through the use of scalable implementations and well-defined loss convergence patterns, the methodology establishes itself as a robust addition to the field of computational linguistics. The proposed approach holds promise for refining the adaptability of language models through iterative knowledge augmentation. Empirical evidence suggests that the refinements foster not only performance consistency but also enhanced domain alignment. The contributions of this work provide a foundation for advancing the operational and representational dimensions of large-scale language models.

We plan seek triplets corresponding y3-plet within n-DJC groups n D. A n D has three constituents, its DJC, removable conflict matrix L and rotation matrix L. In this paper, we will discuss the calculation of removable conflict matrix L which is given by crossed union of JEDS in different DJC of n D. Moreover, the configurations with crossed conflicts and unreasonable conflict can be determined to be 3-uncolorable. And the calculation of L of 4 D is given synchronously.

Omnidirectional images (ODIs) capture 360-degree scenes but require extremely high resolution to provide a truly immersive experience. However, most existing ODIs are insufficient in resolution, leading to degraded visual quality. While prior methods like OSRT have made progress in addressing this issue, they still struggle to fully exploit the geometric properties of equirectangular projection (ERP), particularly in handling geometric distortions and preserving fine details. To overcome these challenges, we propose a Multi-scale Fractal Analysis and Self-adaptive Spherical Harmonics (MFASH) method, which builds upon the OSRT framework to enhance the performance of omnidirectional image super-resolution (ODISR). MFASH introduces fractal geometry to adaptively capture varying levels of detail, self-adaptive spherical harmonics for flexible geometric representation, and nonlinear diffusion equations for edge preservation and noise reduction. By integrating these advanced mathematical models into the OSRT architecture, our method better handles the complexities of ODIs, including geometric distortions and texture preservation. Extensive experiments demonstrate that MFASH outperforms state-of-the-art methods in both quantitative metrics and visual quality, especially in reconstructing fine details and complex textures in ODIs.

This research is in the domain of Natural Language Processing and delves into statistical methods that are intrinsic to the input text, that is not needing any external corpus or training. The methods are used to calculate similarity of sentences in Hindi. The study is performed on sentence pairs in English and Hindi and the differences are analysed with the structural and grammar attributes of the languages. The methods chosen mathematically operate at different levels from character to phrase. The purpose is to identify methods at which a language performs differently than the other, and how different they are with the actual truth. The conclusions can be elemental when considering building larger applications that may need to use one of these methods or hybrid of these methods. An analysis on various methods applied on a sentence pair dataset with both the languages Hindi and English is conducted using visualisation of the produced output of these methods against their actual scores.

This research characterizes properties and some computational features of one big sparse matrix, its condition number, convergence according to the CG method, distribution of nonzero entries, and finally the general sparsity pattern. The condition number factor provides grounds for the numerical stability and accuracy of the solutions concentrated around an eigenvalue magnitude of 0.4. It shows that this matrix is well-conditioned for computational methods as the most important eigenvalues fall into quite a stable range. Now the performance of the Conjugate Gradient method has been considered by monitoring the relative error against iterations. The results obtained show the smooth convergence of the method to an optimum solution, where the relative error decreases steadily with increased iterations, hence proving the efficiency of the method in solving sparse systems. Moreover, the distribution of the values of the non-zero elements has a dominant concentration of around 0.5, which states that most of the non-zero entries of the matrix are around this value. This property is important in numerical analysis of the matrix properties, as well as the opportunities for optimization when performing solutions to systems of linear equations. Lastly, the sparsity pattern of the matrix reinforces the motivation of exploiting sparse matrix techniques, where a large portion of the elements are zero with a sprinkle of non-zero values across it allowing significant savings in memory and computations while employing sparse matrix solvers. Taken together, these results underscore the structure and solvability of large sparse systems and offer a glimpse into the reasons for the performance of numerical methods for their solution.

Deep Neural Networks (DNNs) have revolutionized fields such as computer vision, natural language processing, and speech recognition. Despite their impressive performance, the high computational and memory demands of DNNs present significant challenges for deployment on resource-constrained devices, such as mobile phones and edge computing platforms. Network quantization has emerged as a promising solution to address these challenges by reducing the numerical precision of DNN weights, activations, and gradients, thereby achieving substantial reductions in model size, computation requirements, and energy consumption. This paper provides a comprehensive survey of state-of-the-art network quantization techniques for DNN compression. We categorize these techniques into uniform, non-uniform, and adaptive approaches, analyzing their theoretical foundations, practical implementations, and hardware considerations. Key evaluation metrics, including accuracy retention, computational efficiency, and energy savings, are discussed to highlight the trade-offs involved in applying quantization. Furthermore, we explore recent advancements, such as quantizationaware training, post-training quantization, and hybrid strategies, which aim to enhance the scalability and effectiveness of quantized models. In addition to presenting the current state of the art, this paper identifies critical challenges in the field, such as accuracy degradation, hardware compatibility, and scalability to larger and more complex models. We also outline future research directions, including the integration of neural architecture search, dynamic quantization methods, and innovations in hardware design to optimize support for quantized models. By bridging theoretical insights with practical applications, this survey aims to guide researchers and practitioners in advancing efficient, scalable, and deployable DNN solutions.

API fuzzing, a technique widely used to uncover vulnerabilities in web applications, poses significant security risks when exploited maliciously, leading to service disruptions and data breaches. While firewalls can block unauthorized fuzzing attempts, they limit defenders' ability to gather data on attacker methodologies, reducing actionable cyber threat intelligence. Identifying the responsible fuzzers enables defenders to trace the attacker, uncover their motives, and assess the potential impact, which helps security teams prepare more effectively, mitigate attacks, and develop targeted countermeasures to enhance the security of web APIs. However, analyzing the payloads generated by fuzzers remains largely unexplored and presents significant challenges. For instance, fuzzers often generate similar payloads due to shared initial seeds and similar fuzzing strategies, making accurate fuzzer identification more complex. To analyze this, we experimented with four well-known API fuzzers; APIFuzzer, Kiterunner, RESTler, and Schemathesis, and created a comprehensive dataset of their payloads targeting five different web APIs. Our thorough analysis reveals that the overlapping payloads, i.e., the identical generated payloads across these fuzzers, can be substantially large. For instance, ≈17% of payloads generated with Schemathesis overlapped with ≈12% of the payloads generated with RESTler across different web APIs. As a result, defining distinctive payload features that machine learning models can learn to differentiate and identify their fuzzer accurately becomes more difficult. Alternatively, deep learning techniques, known for their ability to automatically extract features, present a compelling alternative. To evaluate this, we experimented with an architecture combining a bidirectional Transformers-based encoder-decoder and a machine learning classifier to classify fuzzers based on their payloads. Rigorous evaluation using k-fold cross-validation demonstrated high precision and recall, averaging 89%, showcasing this combinatorial architecture's robustness and effectiveness. Our findings demonstrate the potential of combining deep learning and machine learning for fuzzer identification and enhancing web API security.

This study aimed to explore the use of Foundation Models (FM) for skeletal muscle segmentation from magnetic resonance (MR) images, which is a critical step in extracting morphological and functional biomarkers in neuromuscular disorders (NMDs). The research addresses challenges such as fat infiltration and ambiguous muscle boundaries by comparing the performance of traditional convolutional neural networks (CNNs) and emerging FMs like the Segment Anything Model (SAM) and its medical adaptation, MedSAM. A dataset of MR thigh images from 76 NMD patients, splitted in Early, Moderate, and Severe according to the degree of fat infiltration, was annotated for 12 muscle groups across 152 volumes. Fine-tuned SAM and MedSAM configurations (encoder-decoder and decoderonly) were evaluated alongside 2D and 3D nnU-Net CNN models. Performance was assessed using Dice Similarity Coefficient (DSC), Average Symmetric Surface Distance (ASSD), and 95% Hausdorff Distance (HD95). Additionally, uncertainty quantification (UQ) metrics, including Expected Calibration Error (ECE) and Negative Log-Likelihood (NLL), were employed to assess model reliability. SAM's finetuned encoder-decoder configuration achieved segmentation accuracy comparable to state-of-the-art 3D nnU-Net models (DSC: 0.925, 0.883, and 0.857 for Early, Moderate, and Severe cases) and significantly outperformed 2D nnU-Net model. SAM achieved superior calibration, particularly in the Severe group (ECE: 0.032 vs. 0.071 of nnU-Net 3D). Deep ensemble methods further improved segmentation reliability and UQ assessment. MedSAM did not surpass the SAM performance. In conclusion, SAM has demonstrated accurate and better calibrated segmentation, demonstrating its potential for medical imaging applications in challenging NMD datasets.