The increasing sophistication of cyber threats demands advanced and transparent intrusion detection systems (IDS) for the Internet of Medical Things (IoMT). While deep learning has enhanced IDS performance in handling complex data, its black box nature and lack of interpretability undermine trust, accountability, and forensic analysis in critical healthcare environments. This paper proposes XBiLD-IDS, a novel explainable IDS framework designed specifically for IoMT networks. It introduces a transparent feature selection mechanism that integrates SHapley Additive exPlanations (SHAP) with human expertise, enabling interpretable and reliable model insights. This explainable layer is embedded within a hybrid deep learning model that integrates bidirectional long short-term memory (BiLSTM) network with a deep neural network (DNN), which effectively captures temporal dependencies and high-level feature interactions. By quantifying each features contribution to model predictions, our XBiLD-IDS allows experts to refine the feature space for both efficiency and interpretability. Evaluation on the CICIoMT2024 dataset demonstrates the framework’s robustness, achieving 98.70
Coordinating multi-UAV path planning in dynamic environments suffers from severe online computational bottlenecks and non-stationarity. To address these challenges, we propose a physics-prior-driven decentralized deep reinforcement learning (DRL) framework. Functioning as a scalable distributed computing paradigm via decentralized training with decentralized execution (DTDE), the framework mitigates the curse of dimensionality. First, an improved artificial potential field (APF) translates three-dimensional (3D) spatial forces into deterministic control priors, continuously blended with a multi-agent soft actor–critic (MASAC) policy to accelerate offline convergence and save high-performance computing (HPC) resources. Second, embedded long short-term memory (LSTM) networks process locally augmented states for implicit neighbor intent inference without communication overhead. Finally, by integrating these components into the unified APF-LSTM-MASAC architecture, we shift computational burdens to offline HPC training, enabling real-time online inference at the edge. Compared to baseline DRL algorithms, it significantly improves trajectory smoothness and reduces energy consumption by 15.04
For low-latency applications, containerized service deployment at the network edge is essential, but it is limited by a fundamental trade-off between ensuring performance through appropriate computational resource allocation and minimizing storage costs through image layer sharing. Current methods frequently fall short of tackling this joint optimization problem holistically, finding it difficult to manage the hybrid discrete-continuous action space, dynamic network congestion, or the intricate relationship between physical topology and service layer dependencies. In this paper, we present the Synergistic Policy network for Edge Resource Control (SPERC), a novel deep reinforcement learning framework designed to solve this joint problem. SPERC features three key advancements: (1) a dual-stream state encoder combining a Graph Convolutional Network (GCN) for physical topology modeling and a Factorization Machine (FM) to capture latent service-layer dependencies, (2) a differentiable hierarchical control policy where continuous resource allocations from a DDPG agent are seamlessly integrated into the advantage stream of a Dueling Deep Q-Network (DQN) for discrete node selection, and (3) a preference-conditioned learning mechanism enabling zero-shot adaptation across arbitrary latency-versus-storage trade-off preferences. Extensive experiments across edge topologies up to 100 nodes demonstrate that SPERC significantly outperforms state-of-the-art baselines—including GAHRL and Parameterized DQN—achieving higher net revenue, superior resource acceptance ratios, and robust scalability under high service loads.
Accurate indoor localization of unknown radio emitters is important for wireless communication security. Received signal strength difference (RSSD) reduces dependence on unknown transmission parameters, but RSSD fingerprints may contain noise, redundancy, and ambiguous spatial matches. This paper proposes an adaptive singular value decomposition, convolutional neural network, and grey correlation degree (ASVD-CNN-GCD) framework. ASVD projects RSSD fingerprints into a denoised low dimensional subspace, and CNN extracts discriminative features for GCD based similarity matching. A density based spatial screening stage then selects spatially consistent candidate reference points before weighted coordinate estimation. The framework separates offline model construction from online localization, while fingerprint similarity calculations can be executed in parallel using multicore CPUs or GPUs. Simulations and physical experiments demonstrate improved localization accuracy over the comparison methods, with an experimental MAE of 0.5013 m and an average online latency of 7.7083 ms per query.
Telemetry data, essential for operational intelligence and security, is produced in vast quantities by modern cloud data centers. In large-scale enterprise data centers operating hundreds of 100/400 Gbps links, this telemetry stream reaches 50 to 100 Terabytes per rack daily (exceeding 2 to 5 Petabytes per day facility-wide across millions of concurrent flows). This deluge of packets is processed by endpoint server CPUs in traditional storage architectures, resulting in a significant performance bottleneck that restricts visibility. Current methods ignore the network fabric’s computational potential for data orchestration, treating it as a passive collection of pipes. We introduce Synapse, a new switch-centric architecture that turns the network into an orchestrator for active storage. Synapse incorporates intelligent storage logic directly into a programmable switch’s data plane. Bypassing host CPUs entirely, the switch uses Remote Direct Memory Access (RDMA) over RDMA over Converged Ethernet (RoCE) to directly control data placement on a pool of distant Non-Volatile Memory Express (NVMe) SSDs. This design uses policy-driven Quality of Service to safeguard important data streams, autonomous data-plane resilience for microsecond-scale failover and rerouting, and zero-CPU-overhead in-fabric live indexing to make telemetry data query-ready upon ingress. Our analysis demonstrates that in-fabric live indexing accelerates diagnostic queries by up to 300x compared to unindexed scans, while eliminating the severe host CPU overhead (consuming over 90
Multi-modal knowledge graph completion (MMKGC) predicts missing facts by integrating structured triples with heterogeneous visual and textual information. Large-scale MMKGC is also computation-intensive because fine-grained token encoding, expert transformation, contrastive optimization, and candidate-entity scoring must be repeatedly performed over high-dimensional representations. However, conventional MoE routing and contrastive learning remain insufficient when used side by side. Smooth dense routing may mix noisy heterogeneous signals across multiple expert outputs and weaken specialization, leading to what we term “gradient blurring,” whereas overly sharp contrastive optimization may excessively separate structurally related entities and distort local neighborhoods, a potential effect termed “topology tearing.” Routing controls how features are transformed but not their inter-entity geometry, while contrastive learning shapes that geometry without controlling expert allocation. To bridge this gap, we propose TopoMoE, a topology-fidelity discrete mixture-of-experts framework that coordinates input purification, sparse routing, contrastive optimization, and algebraic scoring. Modality-aware gating and an asymmetric information bottleneck first suppress unreliable modality-specific signals. Gumbel-Softmax exploration followed by discrete Top-K selection then assigns each token to selected expert pathways, promoting differentiated feature processing. Lower-bounded adaptive temperatures and curriculum-driven hybrid hard-negative mining subsequently regulate the sharpness and sample composition of contrastive supervision. Finally, a zero-MLP TuckER scorer directly evaluates the optimized representations without additional MLP-based projection layers. This composition connects cleaner routing inputs and differentiated expert processing with bounded contrastive optimization, while retaining batched tensor operations suitable for accelerator execution. Across five independent runs, TopoMoE achieves an MRR of 39.18 ± 0.21% and a Hit@1 of 31.36 ± 0.24% on DB15K, and an MRR of 37.78 ± 0.11% and a Hit@1 of 31.80 ± 0.19% on MKG-W. Supplementary reference-value t-tests yield p<0.01 across the evaluated metrics. Direct Topological Neighborhood Preservation analysis shows the clearest fusion-related improvement in the DB15K Degree >30 group, where TNP increases from 12.90% to 19.80% . In the controlled single-GPU efficiency comparison, TopoMoE records 155.01 G estimated FLOPs and 16.02 s per epoch, compared with 228.24 G and 17.01 s for the dense baseline. These results support a favorable accuracy–computation trade-off and topology-preservation behavior, particularly in dense structural regions, within the evaluated MMKGC settings.
Security is a critical concern in portable devices, which often store sensitive personal and business information. Therefore, it is essential to implement robust protection mechanisms to safeguard against data breaches and cyberattacks. The Advanced Encryption Standard (AES) provides a strong layer of protection that meets global standards. Its symmetric encryption mechanism enables light encryption and decryption processes, making it practically suitable for portable devices with limited battery life. Additionally, the hardware implementation of cryptographic algorithms is more efficient and faster than software solutions, but encounters significant challenges. This paper introduces PipSORE, a novel architectural optimization focused on area and power. The proposed PipSORE has a one-round shared key expansion. The single- and multi-core PipSORE architectures are proposed and evaluated. The architectures also utilize the outer-round pipelining technique. The designs are synthesized and implemented with the Virtex UltraScale VCU108 FPGA board and ASIC NanGate 45 nm technology. The ASIC implementation of the proposed key expansion block achieves a 35
Advanced-node System-on-Chip (SoC) floorplanning is an NP-hard multi-objective optimization problem involving placement legality, half-perimeter wirelength (HPWL), power density distribution, thermal behavior, and timing robustness under fixed-outline constraints. This paper presents QRL-FP, a feedback-aware hybrid framework that combines Quantum-Inspired Optimization (QIO) for coarse global exploration with Deep Reinforcement Learning (DRL) for legality-preserving local refinement. QRL-FP encodes macro-to-bin placement decisions as a feedback-weighted Quadratic Unconstrained Binary Optimization (QUBO) model and applies a Quantum-Inspired Evolutionary Algorithm (QIEA) to generate Pareto-diverse seed floorplans. These seeds are refined using DRL actions guided by multi-channel spatial objective maps. An Adaptive Feedback Loop (AFL) reweights QUBO penalties and DRL rewards according to objective-wise performance gaps, improving coordination under conflicting metrics. Integrated into an OpenROAD-based flow with HotSpot-style thermal analysis, QRL-FP is evaluated on synthetic macro-block benchmarks, Rocket Chip–class SoCs, and FloorSet constrained designs. Across the evaluated benchmarks, QRL-FP reduces HPWL by 15.0–29.6
Deploying high-fidelity deep learning models for robotic harvesting on edge platforms creates a fundamental conflict: the inference latency often exceeds stringent real-time control deadlines, resulting in the critical "stale data" problem. We present a cohesive edge computing system that resolves this conflict through two core contributions. First, we introduce M2PE, a high-fidelity perception engine that achieves an accuracy of 97.4
Metaheuristic algorithms have gained significant attention due to their ability to approximate global optima in complex optimization problems. In this study, a novel nature-inspired optimization method, termite thermoregulation optimization (TTO), is proposed based on the thermoregulation behavior of termites during nest construction. The algorithm models the dynamic balance between exploration and exploitation through a unified adaptive mechanism controlled by the Termite Colony Index, which integrates population diversity and convergence progress. TTO further incorporates directional derivative estimation, bioconvection-inspired oscillatory dynamics, and stochastic diffusion to enhance search capability. The proposed method is evaluated under a unified and fair experimental protocol using identical function-evaluation budgets across all algorithms. Performance is assessed on continuous benchmark functions (Sphere, Rosenbrock, Ackley, Griewank, Schwefel) and discrete Traveling Salesperson Problem instances from the Traveling Salesperson Problem Library, including berlin52, eil51, and kroA100 using a representation-level adaptation based on random-key encoding. Statistical analyses, including multiple independent runs and non-parametric tests, are conducted to ensure robustness and reliability. Experimental results demonstrate that TTO achieves competitive performance compared with established methods such as PSO, ACO, GA, SA, and LSHADE, although its relative convergence accuracy varies across problem types and dimensional settings. Because each TTO iteration performs an independent directional probe and candidate evaluation for every population member, its computational workload increases with population size and becomes particularly important when objective evaluation involves high-dimensional models, engineering simulations, or other computationally expensive processes. The individual-level independence of these operations provides a direct basis for parallel evaluation. In addition, the parallel implementation of TTO is investigated within a high-performance computing framework. The parallel results demonstrate workload-dependent behavior. Moderate speedup was obtained under sufficiently large population sizes, while parallel efficiency decreased as the number of worker processes increased. The identical objective values obtained across worker configurations confirm that the process-based implementation preserves solution quality. These findings position TTO as a parallelizable optimization framework for continuous and combinatorial problems, while its current multicore scalability remains limited by process-management and communication overhead.
Metaheuristic algorithms inspired by natural and biological behaviors have demonstrated strong performance in solving complex optimization problems. However, many existing algorithms are primarily designed for single-objective optimization and, therefore, cannot directly address problems involving multiple conflicting objectives. This paper proposes the Multi-Objective Narwhal Optimizer (MONO), a Pareto-based extension of the recently developed Narwhal Optimizer for solving multi-objective optimization problems. The proposed MONO incorporates Pareto dominance, external archive management, adaptive multi-leader guidance, and crowding-distance-based diversity preservation to effectively balance convergence and exploration throughout the search process. The performance of MONO was evaluated on the CEC benchmark suite and compared with eight representative multi-objective optimization algorithms: NSGA-II, NSGA-III, RVEA, IBEA, SPEA2, MOGWO, MOPSO, and MOEA/D. Experimental evaluation was conducted using three widely adopted performance indicators: Inverted Generational Distance (IGD), Spacing (SP), and Maximum Spread (MS). The results demonstrate that MONO achieves improvements of up to 5.8
Connectivity is a fundamental measure of the fault tolerance of interconnection networks, but traditional vertex-fault models do not describe the structural correlation among nearby failures. To overcome this shortcoming, cluster connectivity and super cluster connectivity have been introduced as more refined metrics for evaluating network fault tolerance under cluster fault scenarios. This paper investigates the cluster connectivity and super cluster connectivity of the complete cubic network, a hypercube-based hierarchical interconnection network. Simulation experiments are conducted to validate the fault tolerance and communication reliability of the complete cubic network, demonstrating that the cluster connectivity metrics provide a more appropriate characterization of network resilience under cluster fault scenarios than traditional measures.
With the rapid evolution of cloud computing, heterogeneous hardware configurations have become increasingly common in modern data centers. However, most existing task scheduling approaches assume homogeneous resources and fail to capture the performance differences among heterogeneous CPUs and memory devices of different generations, resulting in inefficient resource utilization and prolonged task execution time. To address these challenges, this paper proposes a memory sensitive Q-learning (MSQL)-based task scheduling method for heterogeneous cloud environments. By incorporating fine-grained modeling of CPU and memory resources, the proposed approach dynamically optimizes task-resource mapping according to memory bandwidth, latency, and capacity characteristics, thereby improving resource utilization and minimizing task completion time (makespan). Specifically, a memory performance scoring model is first developed to comprehensively characterize memory bandwidth, capacity, and latency, enabling accurate estimation of task execution efficiency under different hardware configurations. Furthermore, a memory efficiency factor is integrated into the reward mechanism to enhance memory resource scheduling. By carefully designing the state space, action space, and reward function, MSQL maintains an effective balance between exploration and exploitation, allowing adaptive scheduling decisions under dynamically changing workloads and resource configurations. This mechanism not only shortens task execution time but also alleviates resource contention and improves overall scheduling efficiency. Extensive experiments demonstrate that MSQL consistently achieves superior performance in computation-intensive, data-intensive, and mixed workloads. Compared with conventional first-come-first-served, ant colony optimization, particle swarm optimization, and genetic algorithm-based schedulers, the proposed method significantly reduces makespan while improving CPU and memory utilization, particularly under memory-intensive scenarios. Owing to its adaptive resource allocation mechanism and memory-aware scheduling strategy, MSQL provides a robust and efficient solution for intelligent task scheduling in heterogeneous cloud environments.
Multi-party digital signatures are fundamental for trust in distributed systems, yet existing solutions suffer from quadratic communication overhead and vulnerability to quantum attacks. This paper presents a practical chain collaborative signing scheme based on NTRU lattice cryptography that achieves linear communication complexity and post-quantum security. The core innovation is a sequential chain propagation protocol that replaces all-to-all broadcast with unidirectional neighbor interactions. Each signer obtains an identity-based NTRU secret key via trapdoor sampling, enabling efficient authentication with minimal storage. A lightweight two-step verification process supports asynchronous third-party validation. The scheme is proved existentially unforgeable under chosen-message attacks in the random oracle model, assuming the hardness of the NTRU Shortest Vector Problem. Extensive theoretical analysis shows linear scaling to thousands of participants with significantly lower communication overhead than broadcast-based protocols. The scheme is particularly suited for sequential authorization scenarios such as blockchain smart contracts, supply chain provenance, and IoT attestation chains. This work bridges the gap between multi-party authentication and post-quantum cryptography by introducing a chain protocol that simultaneously achieves linear message growth, identity-based key management, and provable existential unforgeability, providing a practical solution for resource-constrained environments requiring ordered collaborative signing.
Real-time anomaly detection in evolving data streams remains challenging because of continuous concept drift and transient noise. These factors often cause cluster fragmentation and increase false alarm rates. This paper proposes an Adaptive Micro-cluster Flow-Based Anomaly Detection (MF2AD) framework for evolving data streams. MF2AD constructs a multi-granularity adaptive sliding window mechanism to maintain robust baseline statistics at different time scales. MF2AD extends the mature-tentative micro-cluster organization of existing streaming methods by incorporating pool state into sample-level anomaly scoring and online decision-making. The Hierarchical Anomaly Scoring via Dual-Pool Evolution (HDPES) algorithm combines local micro-cluster information with global statistics, while Dynamic Threshold Calculation via Quartile (DTCQ) dynamically calibrates score- and weight-based decision boundaries. Extensive comparative experiments with 13 baseline methods on two synthetic and eight real-world datasets demonstrate that MF2AD achieves the highest average F1-score and ROC-AUC and ranks third in average PR-AUC. Additional concept-drift, runtime, memory, and parallel scalability experiments demonstrate competitive recovery performance and the feasibility of real-time multicore processing of independent streams, providing a basis for distributed deployment.
In industrial control systems, protocols such as Modbus TCP lack a native security mechanism. Although traditional encryption methods have been employed, the fixed function code structure, periodic polling timing sequence and register access patterns can still leak the critical process semantics. The attackers can bypass the cryptographic protection mechanisms and carry out traffic fingerprinting, program reverse engineering analysis, and data exfiltration attacks. We propose an enhanced Modbus communication encryption method named ModArmor-4D. Our key innovation is the establishment of a four-dimensional integrated security loop that combines entropy source quality assessment, device fingerprint authentication, function code obfuscation, and SM4 encryption. This enables endogenous protection at the protocol layer for key generation, identity authentication, semantic hiding, and data encryption. The experimental results showed that the ModArmor-4D model provided multi-dimensional security protection capabilities close to TLS 1.3 under lightweight deployment conditions. The communication time increased by approximately 57.1
Count data in fields namely behavioral ecology, health, economics, and social sciences frequently exhibit excess zeros, significant overdispersion, and heavy tails. Standard approaches like the zero-inflated Poisson and zero-inflated negative binomial models frequently fail to capture such extreme variability, leading to biased estimates and poor model fit. We suggest a new regression model called the zero-inflated Poisson-XGamma model, which utilizes a compound Poisson-XGamma distribution as a flexible for modeling overdispersed and heavy-tailed count data. We employ a hybrid estimation approach combining a genetic algorithm with the BFGS quasi-Newton method, which provides a robust heuristic for locating the global maximum to deal with the optimization challenges inherent in complex mixture likelihoods. Then, we assess the introduced model via extensive Monte Carlo simulations, demonstrating its competitive performance and low bias in parameter recovery compared to standard competitors, particularly under high overdispersion. The simulation study is designed with 24 scenarios and 10,000 replications each, and it is implemented using parallel computing to achieve efficient execution. Also, we validate the practical utility of the model via a real-world data example, where the ZIPXG model consistently outperforms standard zero-inflated and hurdle models, achieving the lowest AIC and BIC statistics. These results recommend that the ZIPXG model provides a parsimonious and powerful tool for researchers dealing with complex, zero-inflated count data.
Harvested wireless body area networks are efficient solutions for medical care systems, where sensors must continuously monitor patients and maintain stable operation by harvesting energy from environmental sources. Indoor lighting is a key energy source; however, its intensity varies spatially and temporally, sometimes resulting in no harvestable energy. Accurate prediction of harvested energy at each sensor is therefore crucial for system stability. This paper proposes a Transformer-based approach, energy prediction using Transformer (EPT), for indoor energy forecasting. To reduce model complexity, the decoder layer is replaced with a lightweight multi-layer perceptron, while retaining the self-attention mechanism to capture temporal dependencies effectively. Experimental evaluation on a publicly available indoor energy harvesting dataset demonstrates that EPT achieves lower prediction errors than conventional baseline methods, with average root mean square error, mean absolute error, and mean absolute percentage error values of 0.34, 0.09, and 0.64, respectively. The results highlight the model's improved accuracy and computational efficiency.
Within the domain of complex cluster optimization, high-dimensional decision-making environments present formidable computational bottlenecks that persistently challenge traditional processing paradigms. To address the need for highly efficient in computationally intensive environments, this paper proposes a novel multi-agent optimization algorithm predicated upon a state-dependent dynamic evolution mechanism, designated as the multi-agent optimizer (MultiAO). The primary contributions of this work are threefold. First, we introduce a computationally efficient algorithmic framework characterized by three distinct evolutionary phases: gravitational accretion and centroid rebellion, orthogonal singularity decoupling, and a hysteresis effect coupled with a memory potential field. These mechanisms accurately capture agent state changes and significantly accelerate convergence across large-scale search spaces. Second, we provide rigorous empirical validation across two dimensions of the CEC2022 benchmark suite to demonstrate its computational superiority. In comparative analyses against five highly cited and five recently proposed algorithms, MultiAO outperformed all compared algorithms, achieving the best average Friedman ranking scores of 2 and 2.25, and we successfully deploy the algorithm to solve complex, large-scale real-world engineering challenges, specifically unmanned aerial vehicle (UAV) delivery path planning across urban, mountainous, and composite terrains. In both single-UAV and multi-UAV cluster contexts, MultiAO consistently exhibited optimal convergence and robust computational efficiency. Notably, in complex multi-UAV scenarios, MultiAO reduced the total optimal path cost by up to 22.5
Precise and robust detection of components on printed circuit board assemblies (PCBAs) remains a critical challenge for automated quality control within the electronics industry. Modern high-speed SMT (Surface Mount Technology) production lines require both extreme localization accuracy and massive computing throughput for real-time, low-latency visual stream processing. However, execution efficiency and detection reliability are severely hindered by complex visual conditions, including highly similar components, large-scale variations, and densely distributed small targets. To overcome these challenges, this study proposes PCBA-DETR, an effective detection framework. Specifically, PCBA-DETR incorporates three major innovations. First, to alleviate the imbalance between global and local information caused by large-scale variations, a Hierarchical Selective Path Aggregation Network (HSPAN) is designed to adaptively fuse multi-scale feature maps using a feature selection and contextual anchor attention mechanism. Second, a Multilayer Adaptive Refinement Box (MAR-Box) module is designed to improve localization accuracy during the regression process, particularly for small and complex targets. Third, an efficient multi-scale attention (EMA) mechanism is introduced to establish long-range dependencies and enhance the discriminative capability for densely arranged components. Extensive experiments demonstrate that PCBA-DETR achieves superior performance, yielding AP_50:95 , AP_50 , and AP_75 scores of 64.7
