Composite artificial intelligence (AI) systems are increasingly deployed in mission-critical environments, such as defence, aerospace, industrial control systems, and critical infrastructure, where they enable adaptive, autonomous, and real-time decision-making. However, the growing complexity of these systems introduces multilayered security risks that extend far beyond the assumptions of traditional, component-centric security models. In this work, we introduce a structured taxonomy that decomposes composite AI systems into five tightly interconnected layers: core AI and machine-learning (ML) components, integration and orchestration mechanisms, data flows and shared computational resources, cross-layer system interactions and emergent vulnerabilities, and legacy or deterministic software modules that coexist with AI. Leveraging this taxonomy, we propose a holistic, cross-domain threat modelling approach to systematically identify threats, architectural weaknesses, and design-level security flaws across the entire system lifecycle. Finally, we outline mitigation strategies and architectural best practices aimed at building secure, resilient, and trustworthy composite AI systems capable of operating safely under adversarial conditions.

We identify and analyze the Hybrid Cascade Delegation (HCD) vulnerability. HCD is a complex delegation chain formed by combining NS and CNAME records. A carefully constructed HCD chain can exploit and bypass resolvers’ constraints on the number of outgoing queries per resolution. This vulnerability forces a resolver to issue a large volume of queries, resulting in significant amplification of DNS traffic. Currently employed defense technologies are unable to mitigate the threats posed by this vulnerability. In modern consumer electronic devices where service endpoints are embedded and vendor-controlled, the failure of DNS resolution to the core service infrastructure can easily cause high-intensity and large-scale impact. We investigate the HCD vulnerability and address this critical issue in consumer electronic environments. We also propose and discuss mitigation solutions for the HCD vulnerability.

Medical Cyber-Physical Systems (MCPS) are the prerequisites for a highly digital environment in medical facilities. All the technologies, methods, and problems that arise in developing MCPS should be mitigated systematically and continuously to meet this need. This study analytically discusses the currently available MCPS models, threat scenarios, attacks, and data breach statistics to evaluate the current security situation of the research community in the MCPS domain. The large and complex components of MCPS comprise many computational and physical aspects, making managing secure reliability challenging. To address this challenge, there is a need for security by design that incorporates built-in security functionality in the development of an MCPS. Our comprehensive review critically analyzes the current state-of-the-art in MCPS by considering various aspects, such as functionality, communication media, design characteristics, and security. Furthermore, we explore the security threats posed by MCPS as well as their potential impact. Finally, we propose a potential future pathway to effectively mitigate the identified threats by providing a promising direction for further investigation.

Delivering seamless, low-latency, and high-throughput wireless connectivity is essential for emerging metaverse applications, where users engage in immersive experiences such as VR conferencing, digital classrooms, and virtual exhibitions. These applications demand real-time interaction and high network reliability; otherwise, latency spikes, disconnections, and degraded realism can severely compromise user experience and limit largescale adoption. To address these challenges, we propose a robust optimization framework for the deployment of unmanned aerial vehicles (UAVs) aimed at providing uninterrupted wireless connectivity to metaverse-based users, herein referred to as meta users. The novelty of this work lies in jointly optimizing UAV placement, user association, user-UAV distance, and cost efficiency while incorporating spatial uncertainty and user mobility. We formulate a multi-objective integer linear programming model and develop an efficient hybrid method that uses branch and bound (B&B) algorithm and a low-complexity iterative rounding algorithm (IRA). We specifically target the worst-case scenarios by considering robust user associations, especially for edge and mobile users to ensure scalable and adaptive wireless coverage in a dynamic 3D metaverse space. The results demonstrate that the framework enhances scalability, adaptability, and resilience of 3D wireless coverage, establishing a promising foundation for future exploration into autonomous UAV coordination and intelligent, self-optimizing network infrastructures for next-generation immersive communication systems.

The Internet of Consumer Electronics comprises a vast and diverse ecosystem of smart devices ranging from home appliances to autonomous vehicles that rely on DNS for seamless service discovery. However, the rapid expansion of consumer electronics, along with evolving threats, has made cache misses a severe performance and security bottleneck. A cache miss occurs when a resolver does not find an answer locally and then must forward the query to external servers, increasing delay and network load. Attackers can deliberately exploit this behavior through NXDOMAIN-based attacks, which flood the resolver with queries for non-existent domains, forcing repeated upstream lookups and consuming significant processing and bandwidth resources. To address these issues, we propose NXDOMAIN-Aware DNS (NA-DNS), a lightweight enhancement that uses a compact Bloom filter–based representation of DNS zones, enabling resolvers to locally determine whether a queried domain might exist before sending external requests. Experimental results show that NA-DNS can effectively mitigate NXDOMAIN-based attacks, while maintaining fast and stable name resolution performance compared with existing caching methods.

Connected and Automated Vehicles (CAVs) are increasingly recognized as consumer electronics devices with stringent cybersecurity requirements.Traditional Security Information and Event Management (SIEM) architectures are difficult to deploy in Connected and Automated Vehicles (CAVs) because they must operate under bandwidth constraints, limited on-board resources, and strict latency requirements. This challenge must be addressed to enable effective event aggregation, correlation, and fleet-wide monitoring without overloading the vehicle or delaying security response. In this context, we present a hybrid SIEM architecture for CAVs, combining an on-board lightweight SIEM augmented by remote edge SIEM. The architecture uses a security event distribution strategy that incorporates lightweight AI-based event prioritization directly within the vehicle to filter, rank, and forward only the most relevant security events. We validate a proof of concept implementation of our architecture and we show that, compared to traditional SIEM architectures, in-vehicle processing significantly reduces the number of transmitted events by up to 89% and improves the detection latency by up to four times, while preserving global monitoring and cross-vehicle correlation.

The popularity of smart devices extends emerging intelligent applications to remote areas. The limited computing capacity of devices and inadequate ground computing facilities make it a challenge to efficiently process computation-intensive tasks. Fortunately, satellite edge computing networks can provide powerful computing services for users in remote areas. However, the mobility of satellites and the spatio-temporal characteristics of networks bring great challenges to multi-satellite collaborative computing. To address these challenges, we propose a device-satellite-satellite collaborative edge computing network that jointly optimizes task offloading decision and ratio, bandwidth and computing resource allocation. Since traditional optimization algorithms cannot handle time-varying NP-hard problem, we propose an intelligent task offloading method based on multi-agent twin delayed deep deterministic policy gradient. To address the high algorithm complexity brought by high-dimensional action space, we decompose the joint optimization problem into task offloading subproblem and computing resource allocation subproblem. We combine the Lagrangian optimization method to establish a satellite computing resource allocation mechanism, which is convenient for satellite to quickly allocate its edge computing capability according to task requirements, and reduces the dimension of the agent action space. Finally, we propose an Intelligent Task Offloading and Lagrange Optimization Assisted Resource Allocation (ITO-LOARA) algorithm to maximize the task success rate and minimize the task execution delay and energy consumption. Simulation results show that our proposed scheme achieves efficient task offloading and collaborative computing between ground devices and satellite edge computing nodes, and has superior performance, better scalability, and higher stability compared with the baseline and mainstream algorithms.

Autonomous cars are emerging as a key technology in modern transportation, with growing use in personal mobility and shared ride services. Their safe and reliable operation depends on well-defined standards. We review the current status of autonomous car standards, outlining the roles of major standardization bodies and classifying relevant standards into four main categories, namely those that cover safety, communication, cybersecurity, and data management. We identify reliability-related challenges such as regulatory differences, validation of artificial intelligence (AI)-based functions, inadequate simulation frameworks, evolving cybersecurity risks, and the lack of unified performance metrics. We also identify gaps in existing efforts and discuss research and policy directions needed to achieve global harmonization. The results will provide a clear reference for stakeholders working toward the safe, reliable, interoperable, and scalable deployment of autonomous cars.

Intrusion detection systems play major role in security of Internet of Things (IoT) networks against various types of cyber threats. However, traditional Machine Learning (ML) and Deep Learning (DL) models still struggle with challenges such as critical class imbalance, reliance on manual tuning of hyperparameters, and the expensive cost of obtaining labeled data. In this study, we addressed these limitations using the TON_IoT dataset. To address data imbalance, the proximity weighted random affine shadow sampling technique is utilized. Further, Bayesian optimization is applied on LeNet, resulting in LeBayesNet, which provides the optimal configuration for high-accuracy threat detection. Additionally, EntroLeNet integrates entropy-based uncertainty into the learning process for improved robustness. Next, MargiLeNet leverages marginal-based active learning, annotating the most uncertain samples. Experimental results obtained show that LeBayesNet, MargiLeNet, and EntroLeNet improve performance over existing ML and DL models by 6.90%, 5.80%, and 4.27% in accuracy and 6.19%, 6.43%, 8.05%, and 7.52% in receiver operating characteristic-area under the curve, respectively. The LeBayesNet, MargiLeNet, and EntroLeNet models significantly reduce Hamming loss by 72.84%, 60.49%, and 43.21%, respectively. For robustness and generalizability assurance, 10-fold cross-validation is used during evaluation and validated the statistical significance of performance improvements using the Mann–Whitney U test. Furthermore, the models’ explainability and interpretability is enhanced through Shapley additive explanations and local interpretable model-agnostic explanations, providing insights into feature importance and decision transparency.

The proliferation of unmanned aerial vehicles (UAVs) and advancements in artificial intelligence (AI) have fundamentally reshaped modern conflict, introducing AI-assisted drone-swarm attacks as a significant asymmetric threat. We present an empirical analysis of human control paradigms, namely, human-in-the-loop (HITL), human-on-the-loop (HOTL), and human-out-of-the-loop (HOOTL); in the context of recent, notable AI-assisted drone swarm incidents that occurred in real-world nation-state conflicts. We also present a structural threat model that focuses on threats specific to the core components of a drone swarm architecture. Finally, we propose several mitigation solutions to secure AI-assisted drone swarms against inherent threats.

The rapid adoption of consumer drones is enabling access to advanced aerial capabilities and autonomous drone swarm operations. However, most existing research focuses on one-time authentication, after which an adversary can manipulate navigation data shared among drones for swarm coordination without disrupting the communication session. To address this issue, we propose a risk-estimator-based authentication protocol that enables continuous authentication between drones. The approach uses Long-Short-Term Memory (LSTM) and an autoencoder to yield a risk score based on drone behavioral characteristics. Experimental results demonstrate that the method effectively detects GPS spoofing attacks for the DJI Phantom 4 drone family and enhances secure and reliable authentication between drones.

Purpose Today, anomaly detection is an important area studied across various fields to recognize unusual observations. This paper aims to explore the role of deep learning (DL) in developing effective anomaly detection systems (ADSs). Design/methodology/approach This study considers DL mechanisms and their applications in modeling ADSs, discussing critical elements of such systems. This study defines anomalies, identify their primary characteristics and describe three distinct DL-based anomaly detection domains: medical, image and video and cyber-physical systems. This study proposes new classifications for each domain and examines them in terms of datasets, detection methods and implementations. Findings This study identifies the drawbacks and limitations of current anomaly detection techniques and presents several DL-based taxonomies and performance metrics for detecting anomalies. Research limitations/implications The paper discusses the limitations and challenges in existing DL-based anomaly detection techniques, providing insights into areas that require further research. Practical implications The study provides practical guidelines for developing and implementing DL-based ADSs across various domains. Social implications Improved ADSs can enhance security, safety and efficiency in medical, image and video and cyber-physical systems, benefiting society as a whole. Originality/value This paper offers new classifications for DL-based anomaly detection in medical, image and video and cyber-physical systems. It highlights the disadvantages of current DL approaches and recommends future research directions for improving ADSs.

Connectivity has become integral to various application domains, with the automotive sector as a prime example. Advances in electronics, computing, and telecommunications have driven the evolution of the connected car ecosystem, transforming it into a data-rich environment that enhances road safety, efficiency, and overall mobility. However, the success of this ecosystem depends on seamless, reliable, and resilient communications. We identify key challenges that may affect communications in the connected car environment and discuss solutions that enhance resiliency and robustness. Finally, we propose a multi-layered network architecture that will enhance communication resilience in the connected car ecosystem.

The rapid development of Space-Air-Ground Integrated Network (SAGIN) has given rise to collaborative computing architectures, where satellites, Unmanned Aerial Vehicles (UAVs), and ground terminals cooperate to support ubiquitous coverage and real-time services. However, the heterogeneity of multi-modal tasks poses significant challenges to traditional offloading strategies. Traditional static priority methods cannot solve the dynamic coupling between time-varying network states and task characteristics, resulting in QoS violations and suboptimal resource allocation. To solve these problems, we propose a hierarchical offloading and collaborative computing framework driven by dynamic priority of mobile users’ tasks. First, we establish a multi-dimensional task priority evaluation model and combine the satellite coverage time prediction and the UAV trajectory error to modify the task priority in real time. Second, we design a hierarchical game mechanism based on cloud-edge-device collaboration, where UAVs/satellites and mobile devices make distributed bidding decisions and priority-aware offloading decisions. Then, we propose a Dynamic Priority-driven Hierarchical Game Multi-Agent Proximal Policy Optimization (DP-HG-MAPPO) algorithm to integrate task offloading and resource pricing strategies to achieve optimal trade-offs among delay, energy efficiency, and load balancing. Using multi-dimensional feature modeling, game-reinforcement learning fusion and adaptive threshold mechanism, we achieve efficient matching between multi-type tasks and multi-level resources. The simulation results show that, compared with the existing benchmark algorithms, our proposed method improves the task completion rate, delay satisfaction degree of high priority tasks and energy efficiency by 11.7

Timely and accurate detection of foliar diseases is critical to protect corn production against severe yield losses. To address the limitations of traditional image-only diagnostics, we propose CornSentry, a novel edge-aware framework driven by multimodal data fusion. Since this study relies on a composite-sample scheme pairing public images with sampled environmental distributions, we developed a foundational proof-of-concept architecture (called CornSentry) rather than a field-validated biological study. Our architecture employs a late-fusion strategy which combines high-level visual representations extracted by Convolutional Neural Networks (CNNs) with in-situ environmental time series (temperature, humidity, soil moisture) modeled by Recurrent Neural Networks (RNNs). Furthermore, the framework integrates a Deep Convolutional Generative Adversarial Network (DCGAN) to overcome class imbalance and broaden feature diversity through synthetic data generation. We evaluated CornSentry using a comprehensive dataset combining visual and real-world environmental loT data. The results obtained demonstrate that our proposed multimodal late-fusion approach achieves a detection accuracy exceeding 98%, yielding a statistically significant improvement (p < 0.05) and a 3-4% increase over unimodal models that rely solely on image-based detection.

With the extensive application of extended reality (XR), intelligent terminals, and the Internet of Vehicles (IoV) to remote areas and disaster regions, immersive consumer electronic applications are driving the demand for global deployment of computing power and low-latency services. To address this challenge, we propose a "cloud center-space edge-air edge-ground terminal" four-layer collaborative computing architecture. We design an intent-driven dynamic resource scheduling mechanism, which directly perceives the quality-of-service requirements of applications through "business intent descriptors." We develop an intelligent computing-aware routing protocol and autonomous offloading decision algorithm to achieve intelligent collaboration of tasks at the terminals, edge nodes, and the cloud. Then, we analyze key enabling technologies such as on-orbit edge computing of satellites and integrated communication-sensing-computing. Simulation results show that the proposed system and mechanisms can significantly enhance the user experience in scenarios such as XR, IoV and smart homes and address challenges such as global coverage and resource scheduling. Finally, we explore solutions to address issues such as dynamic spectrum sharing, communication energy consumption, battery life, protocol standardization, mobility management, and actual deployment. These solutions enable the upgrade of next-generation Internet of consumer electronics.

A query in the Domain Name System (DNS) can result in a negative response from an authoritative server. A negative response indicates a non-existent name (NXDOMAIN) in a zone. Although resolvers can cache negative responses, queried domain names associated with malicious traffic typically vary and make cached items useless. This variability effectively bypasses resolver caches, resulting in high volume NXDOMAIN traffic that can overwhelm authoritative servers. To reduce the volume of negative queries between the resolver and the authoritative server, aggressive caching has been widely adopted by resolvers. This technique uses previously cached denial records (a record type that proves name non-existence) to reduce query traffic. However, privacy requirements have led to the design of several denial record variants that are either inefficient or incompatible with aggressive caching. To overcome this limitation, we propose a novel denial record design that utilizes a Bloom filter to embed more information for the resolver. Our design provides stronger privacy than the legacy denial record approach and exhibits better compatibility than other denial record variants in supporting aggressive caching. Experimental results show that our approach reduces traffic between resolvers and authoritative servers under massive NXDOMAIN queries, improves resolver response performance, and requires fewer cache entries to achieve the same negative cache hit rates compared with the legacy denial record technique.

Multi-access Edge Computing (MEC) integrated with the Industrial Internet of Things (IIoT) is vital for intelligent manufacturing and industrial automation because it enables low-latency and high-efficiency task offloading from resource-limited devices to an edge server. However, dynamic wireless channels and stochastic task arrivals introduce significant uncertainties, leading to queuing delays, inefficient resource utilization, and high energy consumption. Moreover, the lack of future system information makes real-time offloading decisions particularly challenging. To address these issues, we construct both task queues and delay-aware virtual queues, and we formulate a stochastic optimization problem for joint task offloading and resource allocation. The objective is to minimize long-term energy consumption while ensuring queue stability and satisfying task deadline constraints. To solve this problem, we propose a novel Lyapunov-guided multi-agent deep reinforcement learning framework (LYMADDPG), which integrates Lyapunov optimization with Multi-Agent Deep Deterministic Policy Gradient (MADDPG). Specifically, we use Lyapunov optimization to transform delay constraints into a virtual queue stability control problem, converting the original long-term problem into a series of per-slot optimizations. Next, we use MADDPG to learn optimal offloading and resource allocation policies in a distributed and adaptive manner. Extensive simulation results demonstrate that our method significantly outperforms baseline algorithms in reducing energy consumption, ensuring queue stability, and meeting task deadlines. These results confirm the practical effectiveness of our approach and highlight its strong potential for real-world deployment in MEC-enabled IIoT systems.