Computer Science and Mathematics

Abstract: Operational technology (OT) analytics increasingly requires operational meaning without granting enterprise systems direct visibility into, or control authority over, industrial control assets. Conventional telemetry patterns move raw tags, historian data, or protocol messages into enterprise environments, creating reusable operational visibility even when channels are encrypted or nominally read-only. This article proposes Ephemeral Semantic Telemetry (EST), a learning-mediated, commandless data exchange model in which approved OT observations are transformed into temporary, task-bound semantic languages. The model is formalized using information-bottleneck, finite-time inference, transferability, reconstruction-risk, and command-expressivity conditions. A synthetic industrial telemetry experiment compares raw telemetry, stable semantic telemetry, and EST across authorized task utility, unauthorized inference, raw-data reconstruction, inter-batch transferability, and semantic leakage proxies. Results indicate that EST preserved high authorized task performance while reducing unauthorized transfer accuracy to chance-level, increasing reconstruction error, and reducing inter-batch transferability. EST is not a replacement for cryptography, segmentation, or one-way transfer; rather, it introduces semantic exposure reduction: sharing approved operational meaning while limiting raw operational visibility and preventing command-capable representation.

Abstract:

Cilium is among the most widely deployed Container Network Interfaces (CNIs), serving as the default CNI in Google Kubernetes Engine. It extends standard Kubernetes NetworkPolicy (KNP) with two additional types—CiliumNetworkPolicy (CNP) and CiliumClusterwideNetworkPolicy (CCNP)—each with distinct semantics. When all three coexist in a cluster, the resulting composition is difficult to reason about formally, leading to misconfiguration and security incidents. Existing verification tools, KANO and VeriKube, address subsets of the problem but share two critical limitations: neither provides a formal denotational semantics that precisely characterizes the three-layer composition, nor a canonical representation enabling policy-equivalence checking with completeness guarantees. We close this gap with three contributions. First, we develop the first formal denotational semantics for Cilium’s three-layer composed policy—KNP (additive), CNP (deny-wins), CCNP (cluster-override)—and prove that the composite function is Hyper-Rectangular Piecewise-Constant (HRPC-like). Second, we construct a canonical Reduced Ordered Interval Decision Diagram (ROIDD) for the composite policy space and prove a canonicity theorem, enabling policy-equivalence checking as structural isomorphism in O(|ROIDD|) time. Third, we develop certified conflict-detection algorithms for shadow, redundancy, and cross-layer conflict anomalies across all three layers with formal proofs of soundness and completeness. Experimental evaluation on synthetic policies confirms zero mismatches between ROIDD evaluation and ground-truth brute-force, ROIDD compression ratios of 5–15× over the unshared decision tree, and low-microsecond (0.56–1.52 µs) per-packet lookup latency regardless of policy size. A native C++ implementation, evaluated on the same policy dataset, reconstructs the identical decision-diagram structure and classifies each packet in under 45 ns—about 30× faster than the Python reference—confirming that sub-microsecond classification is inherent to the algorithm rather than an artifact of the implementation language.

Abstract: Edge API gateways occupy the critical path between clients and distributed backend services, making their throughput, latency, and resource efficiency central concerns for next-generation edge computing deployments. Conventional synchronous, thread-per-request gateway architectures impose a hard ceiling on concurrency due to operating system thread scheduling overhead and memory consumption that grows linearly with active connections. This paper presents a formal design and empirical evaluation of an asynchronous, non-blocking edge API gateway framework built on an event-driven I/O model, a lock-free request pipeline, and a back-pressure-aware routing layer. We derive closed-form expressions for throughput, queuing delay, and resource utilization under the M/M/c queuing model, and validate these against measurements on a commodity edge node (4 vCPU, 8 GB RAM). Benchmarks against a representative synchronous gateway baseline show that the proposed framework achieves 3.8× higher peak throughput (112,400 vs. 29,600 req/s), 74% lower P99 latency (18 ms vs. 69 ms) at 80% load, and 61% lower CPU utilization per 10,000 active connections. Back-pressure propagation latency from overloaded upstream to client-visible flow control is measured at a median of 4.2 ms. These results demonstrate that asynchronous, non-blocking design is a prerequisite — not merely an optimization — for edge API gateways operating at cloud-native scale.

Abstract: Quantum Key Distribution (QKD) is a technology that uses the laws of quantum physics to create secret encryption keys between two people. Its main advantage is that any attempt to spy on the communication automatically disturbs the quantum signals, making eavesdropping detectable. In this work, we introduce a new QKD method called ModPhase-8 (QUEST). Instead of using only a few quantum signal types, our system uses eight carefully designed signal variations created by adjusting the phase between two very short light pulses. These eight signal patterns increase the uncertainty for an eavesdropper and allow more information to be securely encoded in each transmission. On the receiving side, the system adaptively switches between two measurement techniques based on the prevailing channel conditions. This adaptive detection mechanism enhances reliability and helps maintain low error rates even when the communication channel is affected by noise. We provide a mathematical proof showing that the protocol achieves a high level of security, restricting an attacker’s probability of success to less than one in ten billion. Even when practical imperfections such as signal loss, channel noise, or detector inaccuracies are present, the system continues to preserve strong security guarantees. Simulation studies further demonstrate that the proposed protocol maintains low error rates (approximately 1–2% under low-noise conditions, and below 4.1% in standard operating regimes) while efficiently generating secure cryptographic keys across a variety of channel conditions. A rigorous nine-experiment benchmarking study establishes that ModPhase-8 achieves 6.75× noise tolerance, surpassing even the Gaussian-modulated GG02 protocol (4.01×), together with 10× faster finite-key convergence, and exclusive key generation in high-noise regimes where six out of eight benchmark protocols fail entirely. Three additional detector-parameter experiments demonstrate that ModPhase-8 tolerates up to 20.6% higher timing jitter than BB84 (158.8 ps vs. 131.7 ps), sustains key generation at afterpulsing probabilities exceeding 15%, and maintains the highest composite detector figure of merit across six detector-quality axes. The reported secret-key-rate values are computed under a heuristic d-ary entropy approximation applied uniformly to all protocols; a fully rigorous discrete-modulation continuous-variable QKD numerical security analysis, justified here through the explicit construction of the equivalent entanglement-based protocol and source-replacement scheme, is identified as essential follow-up work. Overall, ModPhase-8 offers a practical, scalable, and highly secure framework for next-generation quantum communication networks.

Abstract: Mythos-class frontier AI systems, defined in prior work by five indicators, exhibit discontinuous cyber-operational behavior that classical kill-chain models and artifact-centric taxonomies such as MITRE ATT&CK and ATLAS do not accommodate. The prior reference architecture specifies a four-layer defense and the Mythos-Class Posture Rubric (MCPR), whose runtime tier empirically detects supervisability-evasion signatures—runtime evidence that an operation is evading effective supervision. This manuscript provides the scaffolding on which those signatures cohere and the cross-operation extensions they motivate, across four contributions. First, a relational systems-theoretic model treating the enterprise as three coupled frames (identity, trust, telemetry), with frame-shifts defined by non-locality, non-sequentiality, and observability collapse, and grounded in systems-theoretic precedents (Ashby, Luhmann). Second, a four-class taxonomy of the relational space: presence, privilege, domain, and observability discontinuity. Third, a cross-operation detection matrix with four mechanisms operating on existing telemetry. Fourth, integration extensions routing the new signals through the existing mitigation stack without parallel primitives. A controlled synthetic evaluation of the distributional-drift mechanism over a 23-analyst population yields a full-sample area under the ROC curve of 0.94 (95% CI 0.92–0.95); with the threshold calibrated on a held-out partition, it detects 73% of adversaries (95% CI 65–79%) at a near-5% out-of-sample false-positive rate, degrading gracefully against subtler adversaries. The contribution is scaffolding and cross-operation extension to the prior architecture, not a competing framework.

Abstract: In this paper the design, development, and validation of a set of low power, low-cost sensor systems intended for environmental monitoring and visitor tracking in cultural heritage sites is shown. These systems have been deployed in 5 pilot cases in 4 different countries in Europe inside the European ARGUS project. Two different approaches have been defined in the design: the static systems installed in fixed positions and the dynamic systems installed in a moving robot i.e., a quadruped robot or a drone. The Baltanás pilot site has served as the primary testing platform, enabling accelerated prototyping and iterative improvement before deployment across additional pilot locations. This paper presents the design criteria, system architectures, and performance evaluation of these sensor networks, including thermal–humidity probes, volumetric water content sensors, wind measurement systems, pollution monitors, and two generations of visitor counting devices.

Abstract: Agentic AI systems depend on classical public-key cryptography for agent identity, tool invocation, inter-agent communication, model integrity, and persistent state, exposing them to a cryptographically relevant quantum computer (CRQC) along two axes: confidentiality (harvest-now-decrypt-later) and integrity (harvest-now-forge-later). Existing post-quantum migration guidance addresses static, operator-controlled enterprise estates, while emerging agent-identity work omits post-quantum cryptography entirely; neither treats non-human-identity-dense, runtime-negotiated agentic systems as a distinct migration class. This paper develops a conceptual framework that does. It organizes agentic cryptography into seven migration surfaces and separates each identity into a credential layer (symmetric, operator-held, low-risk) and a trust-anchor layer (the asymmetric roots that underwrite the fleet). These layers scale inversely: a small set of trust anchors carries a forge-later blast radius equal to the population beneath it, so migration effort and forge-later risk rank the work in opposite orders. A migration matrix and a parametric effort-and-risk model formalize this, yielding the core sequencing rule: migrate anchors first. Because agentic adoption is ongoing, it also reframes migration from a finite inventory into a continuously regenerating problem, distinguishing remediation of the installed base from prevention of new classical-cryptographic debt in future deployments. It closes with oversight and procurement implications for federal post-quantum readiness.

Abstract: This article presents the development and experimental validation of a Wi-Fi-based distributed system for monitoring the technical condition of building structures. The proposed approach is based on a hybrid mesh/ad hoc network architecture, in which sensor nodes function as autonomous cyber-physical elements and communicate via IEEE 802.11 without relying on a centralized wired infrastructure. The research includes the design of the structural and functional system architec-ture, the development of a distributed data transmission algorithm, and the imple-mentation of a multi-sensor monitoring platform that integrates distance, magnetom-etry, and environmental sensors. A mathematical model of the network is introduced, and the key communication parameters affecting the system’s operation are analyzed. Experimental validation was conducted using 500 consecutive measurements, and a comparative analysis of wired and wireless data acquisition methods was performed. The evaluation was based on statistical metrics such as the mean absolute error (MAE) and root mean square error (RMSE), as well as time-series analysis. The results show that the wireless communication channel reliably preserves the temporal dynamics of the measured parameters without data loss. The highest accuracy was achieved for distance sensors and the DHT22 temperature channel, while magnetometric sensors and humidity measurements exhibited moderate variability depending on sensor sen-sitivity and transmission conditions. The proposed system demonstrates high stability, scalability, and fault tolerance, confirming the feasibility of using standard Wi-Fi technology for monitoring the tech-nical condition of building structures. The developed architecture can be effectively applied not only in the construction industry but also in related fields, such as envi-ronmental monitoring, energy systems, and smart infrastructure applications.

Abstract: The rapid increase of mobile users and advancement of widely used applications introduce high network demands for low-latency and reliable mobility management in mobile communication networks. However, the traditional handover approaches are rule-based and rely solely on signal strength thresholds with hysteresis margins, which are prone to ping-pong effects and are unable to adapt to dynamic network conditions. Machine Learning (ML) models have been integrated for handover predictions, but their centralized architecture compromises user data privacy, which conflicts with the General Data Protection Regulation (GDPR). These centralized ML approaches also introduce scalability constraints that limit their effectiveness in dense network deployments. To address these challenges, this work proposes a Federated Learning with Software-Defined Mobile Networking (FL-SDMN) framework, a unified approach that integrates federated privacy-preserving learning with centralized network coordination for intelligent handover optimization in 5G and beyond networks. The framework leverages a lightweight federated ExtraTrees ensemble model with weighted tree-based aggregation to preserve data privacy and SDMN to provide global network coordination. It has a three-layer decision pipeline that transforms handover control from a reactive threshold mechanism into a predictive, standards-aligned optimization process. Evaluation of the framework was done with real-world 5G mobility data in terms of decision latency, unnecessary handover reduction, and scalability across diverse network configurations. The findings indicate that the integration of FL, Extra Trees, and SDMN provides a scalable, privacy-preserving, and deployment-ready solution for intelligent mobility management in 5G and beyond networks.

Abstract: Kubernetes has become one of the most important platforms for deploying and managing modern cloud-native applications. Its ability to automate container orchestration, scale services, and support distributed workloads has made it a central technology in enterprise cloud infrastructure. However, as more organizations place multiple users, teams, applications, and business units inside shared Kubernetes clusters, the challenges of multi-tenancy, privacy, and security become increasingly serious. Although Kubernetes provides native mechanisms such as namespaces, Role-Based Access Control (RBAC), network policies, resource quotas, and admission controllers, these mechanisms mainly provide logical separation. In highly regulated, hostile, or mutually distrusting environments, such soft boundaries may not be sufficient. Recent research has therefore explored stronger approaches, including virtual clusters, hardware-based Trusted Execution Environments (TEEs), eBPF-based runtime monitoring, service mesh encryption, formal policy verification, and automated misconfiguration detection. This paper presents a comprehensive review of Kubernetes multi-tenancy, privacy, and security research published between 2021 and 2026. The review is organized into three major pillars: multi-tenant isolation, privacy protection, and cluster security. For each pillar, this paper discusses the technical evolution of the field, summarizes representative studies, and compares the main approaches based on isolation strength, overhead, maturity, automation level, and practical limitations. The paper also identifies cross-cutting primitives, including eBPF, RBAC, network policy, admission control, and formal verification, that appear repeatedly across the three pillars. Finally, the paper discusses open research challenges and highlights future directions for building more secure, privacy-aware, and resource-efficient Kubernetes environments.

Abstract: The Evil Twin attack, which involves creating rogue Wi-Fi access points that impersonate legitimate networks, remains one of the most persistent and adaptive threats in cybersecurity, despite more than two decades having passed since its first public demonstration in 2005. This paper aims to provide a comprehensive analysis of the evolution of this attack, perceived as an “invisible enemy” due to its low detectability and systematic underestimation in incident reports. The study addresses key questions: how the Evil Twin attack has evolved, how its methods and tools have changed, where it currently stands, and where it may be heading in the future. The paper compiles evidence from conference presentations, academic publications, government reports, industry analyses, and media coverage, as well as selected defense mechanisms such as WIPS, WPA3, Protected Management Frames, ETGuard, and the Trusted Wireless Environment framework. An original taxonomy of Evil Twin attack mutations is proposed, along with a ten-stage Kill Chain model ([A]–[J]) mapped onto the MITRE ATT&CK framework, an exposure time metric T_e as a key evolutionary parameter, and models quantifying attack cost-effectiveness and efficiency. The analysis demonstrates that the Evil Twin remains a persistent and adaptive threat, whose effectiveness stems from the combination of technical vulnerabilities, user trust in familiar network names, and the difficulty of unambiguous attribution and classification of incidents.

Abstract: Bit allocation is a core design problem in spatially correlated sensor fields under limited communication resources since per-sensor bit depth determines quantization fidelity and thus the quality of acquired information. We address this problem by formulating bit allocation as a cooperative game whose payoff is given in the criterion of mutual information, and by using Shapley value to quantify each sensor’s contribution; to ensure this formulation scales well in larger networks, we approximate Shapley values via Neyman stratified sampling. We compare Shapley value-based allocation against four heuristic baselines – uniform allocation, greedy allocation, Voronoi-based geometry-aware allocation, and conditional variance-based allocation – with both randomly distributed and clustered deployments, using five complementary metrics: mutual information, global RMSE, boundary RMSE, worst-10% RMSE, and weighted posterior trace. Numerical experiments on sampled random fields show that stratified sampling achieves tight efficiency consistency with reasonable runtime and scales to larger sensor counts. Reconstruction performance is context-dependent: geometry-aware allocation often performs best under tight budgets, particularly on boundary and tail errors, while Shapley value-based allocation yields the best performance in stringent small-scale fields and becomes competitive under high budgets for global and tail errors. Overall, mutual information and weighted posterior trace provide complementary rankings, highlighting trade-offs between information-centric objectives and reconstruction-error objectives under heterogeneous spatial redundancy.

Abstract: Transport Layer Security (TLS) protects most Internet traffic, yet migration to post-quantum TLS (PQ-TLS) depends on more than publishing a new algorithm standard. Despite rapid standardization of the Module-Lattice-Based Key-Encapsulation Mechanism (ML-KEM) and hybrid TLS groups, practitioners lack a structured way to assess whether deployment ecosystems are ready to adopt PQ-TLS. This letter proposes a six-dimensional readiness and adoption framework and a reproducible maturity rubric (levels 0--4). A case study of Windows, Linux, macOS, and Android applies the rubric to cited vendor documentation. The main finding is that cryptographic and protocol readiness are uniformly high across the case study, while platform and application readiness show the widest gap and act as the primary adoption bottlenecks for software that inherits OS TLS APIs. The study reports documented transition signals, not live PQ-TLS negotiation rates.

Abstract: TinyML autoencoder anomaly detection is widely proposed for embedded sensor networks because the autoencoder’s learned latent representation supports downstream signal characterization that pure threshold detectors structurally cannot. However, the standard frozen-threshold architecture relies on a calibration-time frozen reconstruction-error threshold whose validity has not been characterized on signals containing slow envelope drift. We test a Hammad-style multilayer-perceptron autoencoder baseline against the non-stationary noise model previously used to validate the Temporal Spectral Noise-Floor Adaptation (TSNFA) detector [1] on Cortex-M4F-class hardware, in both per-node-trained and shared-pre-trained variants. Across the configuration matrix we find a structural false-alarm-rate floor in the order of one thousand false alarms per hour per node, two to three orders of magnitude above TSNFA on the same input realisations. Sweeping the proportion of training frames containing transient bursts and the threshold coefficient confirms the ceiling is not transient-driven but correlated to drifting noise. We then introduce a hybrid architecture in which a single scalar drift estimate sourced from a TSNFA detector normalizes each frame before the autoencoder receives it, leaving the autoencoder weights and frozen threshold unchanged. The hybrid delivers two quantified findings. First, full suppression of false positives: the false-alarm cluster rate collapses from 17.95 clusters per hour per node (MLP TinyML-Shared baseline) to 0.00 clusters per hour per node (hybrid) at 12 dB SNR on a 50-node network, matching TSNFA on the same input realisations, with network load reduced by a factor of 260. Second, signal classification: the autoencoder's 8-dimensional bottleneck output separates background noise from two synthetic event classes at 96.0 % balanced accuracy in the hybrid, against 83.1 % in the baseline detector under the same drift — a feature the binary TSNFA detector cannot provide. The hybrid is therefore not a replacement for TSNFA on detection alone-TSNFA dominates at roughly 200× lower compute, but it becomes the operationally-preferred deployment path whenever downstream signal characterization is required alongside binary detection.

Abstract: In this paper, we examine how the VTP2 extended persistence timeout policy affects and influences the performance of distance-vector and link-state routing protocols in the ad hoc network of the New Generation of Open Digital Universities (DOUNG). The problem addressed is that conventional SCTP retransmissions lack good performance when losses result from a path break rather than congestion. In classical SCTP, missing acknowledgments may trigger retransmissions even when the loss is caused by a temporary route failure rather than by congestion. The proposed evaluation uses an NS-3-compatible methodology with IEEE 802.11, SCTP, AODV, DSDV, DSR and OLSR under increasing node mobility. Results are organized by protocol to improve figure readability. The reference outputs show that VTP2 improves packet delivery ratio, throughput, end-to-end delay, SCTP retransmissions and energy consumption. The average gains are higher for AODV, DSDV and DSR than for OLSR, confirming that extended persistence is more beneficial to protocols exposed to route discovery, repair and maintenance phases. These results indicate that VTP2 is a relevant cross-layer mechanism for improving quality of service in mobile, heterogeneous and distributed digital-university environments.

Abstract: Autonomous Underwater Vehicles (AUVs) have emerged as an effective solution for data collection in Underwater Wireless Sensor Networks (UWSNs), addressing fundamental limitations of acoustic communication such as limited bandwidth, long propagation delays, and high error rates. By moving close to each sensor node for direct data retrieval, AUVs improve energy balance, extend network lifetime, and enhance coverage flexibility. However, AUV-assisted data collection introduces complex challenges, including trajectory optimization under energy, latency, and coverage constraints, as well as robustness to dynamic ocean environments, intermittent connectivity, and large-scale multi-AUV coordination. This survey presents a systematic review of 56 representative AUV-assisted data gathering protocols (2011–2025) and introduces a unified six-class taxonomy that consolidates fragmented classifications in the literature. Following a structured screening of more than 200 peer-reviewed records, these 56 protocols were selected for detailed taxonomy-based analysis. The proposed taxonomy spans classical trajectory optimization, clustering-based organization, learning-based single-AUV methods, multi-AUV coordination, hybrid communication and quality-aware strategies, and energy and lifetime management. In addition, we provide a comparative analysis across key performance dimensions, including energy efficiency, network lifetime, latency, Age of Information (AoI), delivery reliability, and scalability, highlighting fundamental trade-offs among these metrics. Our analysis reveals a clear shift toward learning-based and AoI-driven approaches, while identifying critical gaps in real-world validation, scalability of multi-AUV systems, and security-aware cross-layer design. Finally, we outline open research challenges and future directions to guide the development of robust, scalable, and deployable AUV-assisted data collection systems for next-generation ocean monitoring applications.

Abstract: Wildlife-vehicle collisions (WVCs) cause approximately 570 human fatalities in Canada per 20-year cohort, with Alberta accounting for 22% of these and incurring an estimated CAD $300,000 per day in direct and indirect costs. Wildlife fencing combined with crossing structures reduces collisions by ~86% on well-instrumented sites but remains economically infeasible across the majority of rural road kilometres, leaving a substantial collision residual. We present a combined sensor network integrating alternating-side radar nodes (10-m spacing baseline), three-axis magnetometers, dynamic message signs, and LoRa-mediated awareness propagation between adjacent radars. System performance is evaluated through a discrete-time Monte Carlo simulation on a 1 km test corridor, incorporating a six-state animal behavioural Markov model with vehicle-threat-dependent decision branching, Intelligent Driver Model vehicle dynamics, and a three-mode contrast that isolates the contributions of sensing, driver alerting, and network coordination. Across 60 independent trials, the integrated system reduces the collision rate per road entry by 47.4% relative to an unmitigated control (Welch's t = 2.82, p < 0.01), and simultaneously increases safe road-crossing throughput by 77% by lowering the perceived vehicle threat that otherwise triggers pre-crossing retreats. Sensitivity sweeps establish a statistically significant equivalent-performance band across 5-20 m alternating radar spacing and across small-to-medium animal classes (fox- through deer-class), with operational robustness against tenfold degradation of baseline sensor sensitivity. A conservative 20 m alternating deployment spacing is recommended to provide engineering margin against range-dependent radar SNR, clutter, and environmental factors not captured in the idealized detection model. The architecture complements existing fence-and-crossing infrastructure at approximately one order of magnitude lower per-kilometre cost.

Abstract: Industrial systems increasingly rely on MQTT-based message streaming to enable automated, data-driven production processes at the network edge. While semantic models such as the SSN/SOSA ontology enable machine-interpretable descriptions of observations and actuations, an explicit model of message transport is rarely considered. Consequently, MQTT-based communication remains opaque, particularly regarding information processing, hindering the semantic analysis of application-specific topic structures and the behavior of transport protocols. To close this gap, this work introduces the revised MQTT4SSN ontology as a key contribution, extending existing semantic models with protocol-aware representations of MQTT entities, control packets, and transport-level interactions. MQTT4SSN enables end-to-end semantic traceability, from sensor observations and actuator controls to the underlying message transmission within distributed systems. Building on this contribution, the MQTT2RDF integration framework incorporates MQTT4SSN as its core to capture live MQTT traffic and represent both payload meaning and transport-level provenance within an RDF knowledge graph. This work presents a novel approach for representing edge computing and information processing over MQTT, addressing two key challenges. First, a semantic topic-naming approach automatically derives MQTT topic hierarchies and payload content structures from observation and actuation semantics. This approach facilitates the setup of edge computing systems and enables context-aware subscription management and structured data formatting, thereby improving interoperability between heterogeneous deployments. Second, transport-level provenance analytics support automated detection, classification, and root cause analysis of malformed MQTT packets and protocol-level errors. The approach provides explainable, traceable information processing through transport provenance, which is essential for safety-critical industrial environments. The contributions are validated through an industrial use case from a production environment, demonstrating its applicability for system monitoring, troubleshooting, and semantic analytics of MQTT-based infrastructures.

Abstract: Cloud computing and mobile edge computing address the growing demand for computing power driven by the rise in data-intensive applications, but they are prone to creating computing silos, resulting in unbalanced resource utilization. To address this issue, the Computing Power Network (CPN) has been introduced to enable the centralized management and scheduling of resources across the entire network. However, task scheduling in the CPN requires joint selection of computation nodes and routing paths, which greatly increases the complexity of scheduling problem. In existing studies, heuristic methods are difficult to satisfy real-time requirements, whereas deep reinforcement learning methods ignore the collaborative optimization of network resources, making them difficult to adapt to complex CPN scenarios. To this end, we propose a task scheduling method for the CPN, called TS-DQNF. First, the method uses the Deep Q-Network (DQN) to determine the computation node for computation task. Then, it introduces a dynamic congestion-aware mechanism to determine the shortest routing path. Finally, it gradually obtains the optimal task scheduling scheme through multiple rounds of alternating iterations. Simulation results show that the TS-DQNF achieves good performance and good convergence performance under different scenarios and scales.

Abstract: Unmannedaerial vehicle (UAV)-assisted millimeter-wave (mmWave) and terahertz (THz) communications are promising enablers of ultra-reliable and low-latency communication in next-generation wireless networks. However, the initial access and beam alignment process remains challenging because highly directional beams must be rapidly aligned in a three-dimensional environment. In this paper, we investigate a risk-aware beam alignment framework for UAV-assisted mmWave/THz systems, where user equipment scans a 3D spherical region to detect UAV base stations. The objective is to jointly minimize the expected cell-search latency and its variance while satisfying detection-failure and link-quality constraints. To solve this non-convex optimization problem efficiently, we employ the Lévy Self-Renewable Flow Direction Algorithm (LSRFDA), which combines Lévy-flight exploration with self-renewal to improve convergence robustness. A unified propagation model is adopted to cover both mmWave and THz regimes by incorporating free-space spreading loss and frequency-dependent molecular absorption. Extensive Monte Carlo simulations compare the proposed approach with Particle Swarm Optimization, Random Search, Reinforcement Learning, and PPO-Lagrangian methods. The results show that LSRFDA achieves lower latency, lower latency variation, more reliable detection, and lower energy consumption across a wide range of UAV densities and coverage radii. These outcomes highlight the effectiveness of risk-aware geometric optimization for fast and dependable initial access in UAV-assisted 5G mmWave and 6G THz networks.