The widespread integration of Artificial Intelligence (AI) in sixth-generation Internet of Things (6G-IoT) applications, introduces significant challenges for ensuring the trustworthy and dependability of AI models. The "black-box" characteristic of numerous Deep Neural Networks (DNNs) creates a notable obstacle for confirming their safety in intricate, ever-changing environments. Consequently, there is a need for extensive testing, requiring the gathering and labeling of a large number of test cases, a process that is both time-intensive and resource-consuming. While previous studies have adapted neuron coverage criteria for steering test case generation in DNNs. Yet, these criteria are white-box measures requiring access to model states and presenting their practical limitations. Conversely, black-box metrics, which focus on outputs, present a more feasible approach. Among these, black-box diversity metrics evaluate model robustness by generating diverse test cases, eliminating the need for internal model details. This paper presents a test case generation framework centered on diversity, known as BDTest. BDTest enhances test adequacy through five stages: (1) Mapping feature vectors extracted from an initial set of seed images onto a low-dimensional manifold utilizing UMAP; (2) Detecting sparse regions using DBSCAN; (3) Sampling key points from these regions via Latin Hypercube Sampling; (4) Reconstructing latent features and generating new images through ICA and GAN inversion; and (5) Measuring the diversity of the generated set using metrics such as the Log-Determinant. Experiments demonstrate that BDTest significantly improves test set diversity and error detection performance, achieving error rates of 59.36%, 59.76%, and 67.03% on VGG19, DenseNet121, and MobileNetV2, respectively, outperforming DeepXplore by an average of 12.43% and DLFuzz by 9.95% across all tested models. When retrained with the generated test cases, the model demonstrated improved accuracy on the original test set, alongside a significant enhancement in accuracy on the natural adversarial test set.

The integration of Large Language Models (LLM) with multimodal agentic AI within the Medical Internet of Things (MIoT) ecosystem is redefining modern healthcare intelligence. This convergence enables continuous patient observation, adaptive clinical decision-making, and context-aware interaction between humans and machines across various biomedical data modalities. Healthcare systems generate a wide range of multimodal data, including textual records such as EHRs, prescriptions, and pathology notes; medical imagery such as CT, MRI, fundus, and radiographs; spoken data from consultations and transcriptions; video streams for rehabilitation and physiotherapy monitoring; and sensor readings such as ECG, SpO \({}_{2}\) , and glucose levels. Conventional unimodal algorithms fall short in interpreting this diversity, whereas LLM-augmented agentic frameworks fuse and reason over these heterogeneous sources, grounding their outputs in medical ontologies and coordinating task-specific agents to enhance real-world clinical workflows. This article presents a comprehensive overview of multimodal agentic AI powered by LLM for MIoT-enabled healthcare systems. Introduces a 6D unified taxonomy that covers multimodal input channels, fusion mechanisms, core LLM reasoning capabilities, agentic coordination models, computational deployment layers, and ethical governance frameworks. To contextualize this taxonomy, the discussion includes a Virtual Hospital case study centered on cancer that demonstrates how multimodal signals such as imaging, genomics, patient dialogues, and clinical updates integrate through intelligent agents to enable personalized diagnosis, automated documentation, home rehabilitation, and rapid intervention in emergencies. The survey also consolidates current progress on datasets, benchmarks, and evaluation protocols for AI in multimodal and agentic healthcare. The survey identifies critical research gaps, such as the lack of longitudinal multimodal datasets, standardized evaluation frameworks for multi-agent reasoning, and reliable methods to assess trustworthiness in clinical AI. Furthermore, it examines security and compliance issues such as adversarial manipulation, data leakage, and accountability across distributed agent networks, and it proposes countermeasures through federated data governance, secure MCP-oriented orchestration, and privacy-aware edge deployment strategies. By situating recent advances within the Virtual Hospital paradigm and oncology workflows, this study provides a systematic foundation for developing scalable, secure, and ethically aligned multimodal agentic systems based on LLMs, guiding the next generation of intelligent MIoT-driven healthcare ecosystems.

Rapid iterations and updates in both software and hardware, along with significant advancements in communication technology, have given rise to the concepts of symbiotic Internet of Things (IoT) and ubiquitous interconnectivity, providing strong evidence for the flourishing development of the Internet of Things. However, the limited resources and computing capabilities, along with the heterogeneity of deployment environments, make symbiotic IoT devices more susceptible to security threats and operational issues. Intermittent failures are especially prevalent in the symbiotic IoT, leading to more significant risks for devices. In this paper, we present an IFDGAT-LSTM (Intermittent Fault Diagnosis Based on Long Short-Term Memory and Graph Attention Network) framework for diagnosing intermittent failures in wireless sensing devices within the symbiotic IoT. The framework is based on a graph neural network and takes into account not only the time series characteristics of symbiotic IoT devices but also their deployment topology. By incorporating both aspects, we achieve more accurate diagnostics of intermittent failures in the symbiotic IoT, thus enhancing its reliability. Firstly, we propose the concept of a quasi-dynamic graph based on the variations in the topology within the symbiotic IoT. Subsequently, we introduce an intermittent failure diagnosis framework that combines a graph neural network to identify intermittent failure nodes within the quasi-dynamic graph. Finally, we performed experiments on the WADI symbiotic IoT dataset to evaluate the performance of our model in diagnosing intermittent failure nodes. We used the precision, recall, and F1 score metrics for assessment. The experimental outcomes show that our proposed model, IFDGAT-LSTM, achieves an Precision of 99.58% in diagnosing intermittent failure nodes. This highlights the strong performance and efficacy of the IFDGAT-LSTM model.

The rapid proliferation of Internet of Things (IoT) systems has introduced complex networks of interconnected devices, computational resources, and web-based communication infrastructure. Privacy protection in IoT data routing is critical to enabling secure deployment of large language models (LLMs) for processing distributed sensor data, user queries, and device-generated content. However, IoT environments inherently involve heterogeneous devices, dynamic network topologies, and resource-constrained nodes, complicating the design of privacy-preserving routing mechanisms that simultaneously ensure reliability across diverse communication layers. To address these challenges, we propose an innovative FtPR (Fault-tolerant Privacy Routing) model based on secure multiparty computing mechanism, which enables secure and efficient data fusion and transmission in IoT networks. FtPR establishes a novel connection between IoT device clusters and data center network architecture AQDNn routers, leveraging the hierarchical architecture of AQDNn to construct completely independent spanning trees (CIST). By exploiting the non-overlapping paths between nodes in distinct CISTs, FtPR achieves fault-tolerant routing while maintaining privacy guarantees. Building on this framework, we introduce a secure multiparty computing mechanism to perturb link weights in the AQDNn. This ensures that link weights across different CISTs adhere to constrained ranges, preventing adversarial inference of routing paths. Each node operates with localized knowledge of its connected link weights, eliminating the need for global network visibility. Consequently, even if malicious actors compromise one or multiple nodes, they cannot reconstruct end-to-end communication paths, thereby preserving route anonymity. Experimental results demonstrate that FtPR improves IoT network performance and security, reducing misclassification rates and marginal release score compared to state-of-the-art methods.

The development of the next generation ubiquitous network puts forward higher requirements for the connection density in the communication network, e.g., massive IoT and UAV swarm, which has led to a lot of research on link management. With the expansion of network scale, the weaknesses of existing algorithms in computing efficiency, performance, and realizability have become prominent. The emerging graph neural network (GNN) provides another way to solve this problem. In this paper, we design a cross-receptive distributed GNN structure from the perspective of communication system, combining measurable index of the actual scene with message passing frame. This new GNN structure and the additional input feature dimension work together to provide richer and more comprehensive information for network training. After the initial deployment of the power decision from GNN, we select some links to shut down and others to reduce their transmit power to further improve system performance and save energy. Simulation results show that our proposed method reaches 83.1% performance of the centralized mechanism. In addition, the discussion on scalability suggests that in order to save training cost, small-scale scenes with the same density can be selected for training in the application of large-scale scenes.

The emergence of new application scenarios has enabled integrated sensing and communication (ISAC) as one of the potential technologies of 6-th generation mobile communication (6G). To meet both communication efficiency and sensing efficiency, a satisfactory ISAC waveform is essential. In this paper, we primarily investigate the impact of different modulation schemes on sensing performance for UAV-to-ground scenarios. Firstly, we analyze the sensing performance differences of modulation schemes in the sensing algorithm based on orthogonal frequency division multiplexing (OFDM) systems. Secondly, we examine the periodic auto-correlation functions (PACFs) with different modulation schemes and modulation orders. We observe that the waveforms modulated by phase shift keying (PSK) exhibit the lower sidelobes compared to waveforms modulated by quadrature amplitude modulation (QAM). Finally, we simulate the probability of detection (Pd) with different modulation schemes for UAV-to-ground scenarios. Numerical results demonstrate that the modulated waveform with constant modulus exhibit superior sensing performance. For the modulated waveform with non-constant modulus, the higher modulation orders result in the poorer sensing performance. This inspires us to change the sensing performance by designing the power spectrum of the modulated waveform with non-constant modulus. This work is helpful for the waveform design and performance analysis of ISAC.

Industrial cyber–physical systems (ICPSs) can play a crucial role in damage assessment during extreme conditions by leveraging their integration of physical infrastructure, sensing capabilities, and advanced analytics. However, due to the wireless sensing devices that are made to operate in ICPS, there is a dire need to address the green routing (energy-efficient) challenges through an optimized solution. In recent times, artificial intelligence (AI) has had a significant impact on wireless sensor networks (WSNs) designed to operate as ICPS components. In this research work, we present AGRIC: AI-based green routing for ICPS. While following the cluster-based routing, the election of cluster head (CH) is executed using our proposed AI-inspired extended spotted hyena Lévy flight optimization (ESHLFO) algorithm. Furthermore, to address the energy hole problem, four energy-unlimited data collection nodes are used around the periphery of the network. The results of the experiment demonstrate the fact AGRIC delivers network longevity and supreme performance in the context of stability time, throughput, and energy left over in the network as important performance indicators.

Network slicing (NS) is a highly promising paradigm in 5G and forthcoming 6G communication networks. NS allows for the customization of multiple logically independent network slices to provide tailored service for vertical applications with diverse quality of service (QoS) requirements. However, current research on NS primarily relies on the traditional modeling methods such as service function chaining (SFC) and task offloading, which have limitations in adapting to the evolving scenarios in 5G/6G networks. To address this, our study introduces one novel Self-adaptive Network Slicing (SNS) modeling method. In this approach, each service is abstracted as multiple SFC replicas originating from diverse access points. Based on the SNS modeling, we investigate a VNF configuration and flow routing (VCFR) problem for service provisioning in a multi-tier system. With the objective of achieving load-balancing with minimal slice operational expenditure, we formulate the VCFR as a mixed-integer linear programming. However, deriving an exact solution via MILP is computationally expensive due to its NP-hardness. To reduce computational complexity, we propose one Load Balancing-considered Community Detection-based Heuristic (LBCD-Heu), our divide and conquer approach, to solve the problem. In LBCD-Heu, we first design a load balancing-considered community detection method to divide the substrate multi-tier network into multiple independent communities. Following this, the MILP is employed in each community to obtain a near-optimal solution. Extensive evaluations justify that LBCD-Heu can effectively reduce the service operational cost and algorithm run-time while ensuring the load balancing of substrate network. Additionally, our results verify that the SNS modeling enables the provision of services at lower expenditures compared with traditional modeling methods.

The outbreak of Coronavirus Disease-2019 (COVID-19) has posed a threat to world health. With the increasing number of people infected, healthcare systems, especially those in developing countries, are bearing tremendous pressure. There is an urgent need for the diagnosis of COVID-19 and the prognosis of inpatients. To alleviate these problems, a data-driven medical assistance system is put forward in this article. Based on two real-world data sets in Wuhan, China, the proposed system integrates data from different sources with tools of machine learning (ML) to predict COVID-19 infected probability of suspected patients in their first visit, and then predict mortality of confirmed cases. Rather than choosing an interpretable algorithm, this system separates the explanations from ML models. It can do help to patient triaging and provide some useful advice for doctors.

The traditional natural gas Internet-of-Things (IoT) system has many problems, such as centralized management of resources, noncirculation of data between stations, insecurity of transaction information or account books, and lack of contract consensus. In order to ensure data security and reliable transaction, this article introduces artificial intelligence (AI) and blockchain technology and constructs an AI-enabled and blockchain-powered natural gas IoT system in a smart city. In this article, the natural gas output prediction model based on temporal pattern attention-based LSTMs (TPA-LSTMs) is used to enable the system to sense the change of natural gas deliverability. In addition, we establish a blockchain-based secure natural gas transaction scheme, which dynamically matches the purchase contract and sale contract to maximize the interests of the buyer and the seller and obtain a transaction contract. The experimental results show that our model can predict the output value of natural gas in real time and select the appropriate transaction matching scheme according to the dynamic demand for sales.

The Internet of Things (IoT) is a group of millions of devices having sensors and actuators linked over wired or wireless channel for data transmission. IoT has grown rapidly over the past decade with more than 25 billion devices expected to be connected by 2020. The volume of data released from these devices will increase many-fold in the years to come. In addition to an increased volume, the IoT devices produces a large amount of data with a number of different modalities having varying data quality defined by its speed in terms of time and position dependency. In such an environment, machine learning (ML) algorithms can play an important role in ensuring security and authorization based on biotechnology, anomalous detection to improve the usability, and security of IoT systems. On the other hand, attackers often view learning algorithms to exploit the vulnerabilities in smart IoT-based systems. Motivated from these, in this article, we propose the security of the IoT devices by detecting spam using ML. To achieve this objective, Spam Detection in IoT using Machine Learning framework is proposed. In this framework, five ML models are evaluated using various metrics with a large collection of inputs features sets. Each model computes a spam score by considering the refined input features. This score depicts the trustworthiness of IoT device under various parameters. REFIT Smart Home data set is used for the validation of proposed technique. The results obtained proves the effectiveness of the proposed scheme in comparison to the other existing schemes.

Security of networking in cyber-physical systems is an important feature in recent computing. Information that comes to the network needs preevaluation. Our solution presented in this article is based on deep learning model developed for network traffic analysis of various Internet of things solutions. At the level of firewall or gateway, information about current connection is gathered for the recurrent neural network. The model evaluates this information and forwards decision back to the firewall to take security actions if needed. In the research, we have tested our solution on two open datasets. The results confirm that our model is very efficient in recognition of potential threats reaching above 99% of accuracy even in a case of reduced number of evaluated networking features.

Today, with the worldwide offer and rapid increment in multimedia applications on the web, the demands of users to get them accessed are also increasing prominently. The users in vehicular environment too expect efficient multimedia streaming while travelling on the road. However, the high mobility of vehicles as well as the limited transmission range of infrastructure components in IP based network provides low performance by offering high delay and additional network overhead. To provide better Quality of Experience (QoE) with high performance, Information Centric Networking (ICN) is blended with vehicular environment. Caching the content inside network nodes is inherent feature of ICN with various associated benefits such as low content retrieval delay, less network traffic, path reduction and so on. However, challenges still exists for caching the content due to resource constrained network environment (such as limited cache capacity, node battery) as well as for secure delivery of cached data. To solve these challenges and to enhance network performance, we propose a cooperative caching scheme in hierarchical network architecture that jointly considers cache location as well as combined content popularity and predicted future rating score while making caching decision. The proposed approach uses two layer hierarchical architecture where nodes in edge layer are divided into clusters. The proposed scheme uses modified Weighted Clustering Algorithms (WCA) for selection of cluster heads which are then used to decide cache location. A probability matrix is used to compute content caching probability which considers both popularity and predicted future rating of content. The proposed approach dynamically predict the user's preferences using non-negative matrix factorization (NMF) - a machine learning technique which eventually provides prediction of future rating. Based on the selection of both cache location and content to cache, the proposed scheme can effectively cache the content in the network. Further, to deal with the secure delivery of cached content, this work supports legitimate user authorization at edge nodes. The performance of the proposed scheme is evaluated in MATLAB parallel computing toolkit. The results prove significant caching improvement in terms of cache hit, hop reduction and average delay using our proposed scheme.

Vehicle-to-vehicle communication assists road-side information exchange granting ease of access and sharing between users. The communication between the vehicles is short-lived due to interference and data congestion in the resource constraint medium. This manuscript introduces a linear adaptive congestion control (LACC) augmenting the benefits of greedy routing and data dissemination model (DDM). LACC focuses on selecting beneficiary vehicle by assessing its end-to-end service capacity and link stability preference. Different from the conventional greedy approach, routing is aided by a linear integer programming module for smart decisions on neighbor selection. The interrupts in data transmission and forwarding due to non-localized vehicles, congested routing paths and paused transmissions are addressed using LACC as a series of linear optimization. This helps to improve the performance of vehicular communication estimated using delay, message delivery, outage, and beacon messages.

Edge Artificial Intelligence (AI) is the latest trend for next-generation computing for data analytics, particularly in predictive edge analytics for high-risk diseases like Parkinson’s Disease (PD). Deep learning learning techniques facilitate edge AI applications for enhanced, real-time handling of data. Dopamine is the cause of Parkinson’s that happens due to the interference of brain cells that produce the substance to regulate the communication of brain cells. The brain cells responsible for generating the dopamine perform adaptation, control, and movement with fluency. Parkinson’s motor symptoms appear on the loss of 60% to 80% of cells, due to the non-production of appropriate dopamine. Recent research found a close connection between the speech impairment and PD. Many researchers have developed a classification algorithm to identify the PD from speech signals. In this article, Adaptive Crow Search Algorithm (ACSA) and Deep Learning (DL)–based optimal feature selection method are introduced. The proposed model is the combination of CROW Search and Deep learning (CROWD) stack sparse autoencoder neural network. Parkinson’s dataset is taken for the experiment from the Irvine dataset repository at the University of California (UCI). In the first phase, dataset cleaning is performed to handle the missing values in the dataset. After that, the proposed ACSA algorithm is employed to find the scrunched feature vector. Furthermore, stack spare autoencoder with seven hidden layers is employed to generate the compressed feature vector. The performance of the proposed CROWD autoencoder model is compared with three feature selection approaches for six supervised classification techniques. The experiment result demonstrates that the performance of the proposed CROWD autoencoder feature selection model has outperformed the benchmarked feature selection techniques: (i) Maximum Relevance (mRMR) (ii) Recursive Feature Elimination (RFE), and (iii) Correlation-based Feature Selection (CFS), to classify Parkinson’s disease. This research has significance in the healthcare sector for the enhancement of classification accuracy up to 0.96%.

This paper addresses the issues of selfishness, limited network resources, and their adverse effects on real-time dissemination of Emergency Warning Messages (EWMs) in modern Autonomous Moving Platforms (AMPs) such as Vehicular Social Networks (VSNs). For this purpose, we propose a social intelligence based identification mechanism to differentiate between a selfish and a cooperative node in the network. Therefore, we devise a crowdsensing based mechanism to calculate a tie-strength value based on several social metrics. Moreover, we design a recursive evolutionary algorithm for each node’s reputation calculation and update. Given that, then we estimate each node’s state-transition probability to select a super-spreader for rapid dissemination. In order to ensure a seamless and reliable dissemination process, we incorporate 5G network structure instead of conventional short range communication which is used in most vehicular networks at present. Finally, we design a real-time dissemination algorithm for EWMs and evaluate its performance in terms of network parameters such as delivery-ratio, delay, hop-count, and message-overhead for varying values of vehicular density, speed, and selfish nodes’ density based on realistic vehicular mobility traces. In addition, we present a comparative analysis of the performance of the proposed scheme with state-of-the-art dissemination schemes in VSNs.

Adversary models have been fundamental to the various cryptographic protocols and methods. However, their use in most of the branches of research in computer science is comparatively restricted, primarily in case of the research in cyberphysical security (e.g., vulnerability studies, position confidentiality). In this article, we propose an energy-aware green adversary model for its use in smart industrial environment through achieving confidentiality. Even though, mutually the hardware and the software parts of cyberphysical systems can be improved to decrease its energy consumption, this article focuses on aspects of conserving position and information confidentiality. On the basis of our findings (assumptions, adversary goals, and capabilities) from the literature, we give some testimonials to help practitioners and researchers working in cyberphysical security. The proposed model that runs on real-time anticipatory position-based query scheduling in order to minimize the communication and computation cost for each query, thus, facilitating energy consumption minimization. Moreover, we calculate the transferring/acceptance slots required for each query to avoid deteriorating slots. The experimental results confirm that the proposed approach can diminish energy consumption up to five times in comparison to existing approaches.

In the context of smart cities and Internet of Things (IoT), there are many trending contents on the social networks that reflect the picture of the community or their interest. In this paper, we propose a model that automatically collect trending social data and analyze them automatically. The model explores trending contents, overall attitude of textual contents and the relationships among the participated users. The analysis of the data collected involves the analysis of the user as well as the community in terms of interest, embedded tags, and the corresponding contents. The model is trained using data collected from Twitter, the famous growing social network, using hashtags, emoticons, geo-tags, and user profiles. We focus in this work on the Arabic contents to visualize the resulted emotions on a real world map.