Spatio-temporal volatility of mobile demand and the surge of computation-intensive, delay-sensitive applications pose significant challenges to cost-aware Mobile Edge Computing (MEC) deployments. This paper investigates a multi-UAV-assisted MEC framework and proposes a hierarchical three-time-scale resource provisioning and allocation scheme. At the large time scale, we optimize long-horizon resource provisioning decisions to jointly minimize service degradation and resource cost. At the medium time scale, a spatio-temporal fluctuation model is established for urban users and a Future-Aware UAV Strategy (FAUS) is introduced to proactively dispatch UAV-mounted servers toward predicted hotspot regions. Meanwhile, we design a semi-competitive and semi-cooperative Stackelberg pricing game between stationary edge servers and UAV clusters to ensure individual rationality and induce effective workload splitting. At the small time scale, we implement a lightweight per-slot routine for price-aware offloading and KKT-based resource allocation for efficient execution. To solve the resulting expensive black-box provisioning optimization, we develop an Environment Switching Evolutionary Optimization (ESEO) algorithm to accelerate convergence and generate diverse Pareto-optimal solutions, enabling flexible choices such as resource-saving MEC and performance-oriented MEC. Extensive evaluations based on a real urban mobility dataset demonstrate that the proposed scheme outperforms benchmark methods in reducing offload failures and delay while providing flexible cost–performance trade-offs.
Service recommender systems typically infer user preferences using observed user-service interactions. However, these datasets are commonly characterized by a long-tail distribution, leading models to favor popular services. Existing debiasing methods commonly mitigate this issue by reweighting popularity-related training signals, regularizing learned representations, or adjusting ranking scores during inference. While such approaches can increase the recommendation opportunities for long-tail items to some extent, they often overlook the heterogeneous preferences of different user groups toward item popularity and fail to fully leverage semantic knowledge from textual information, thereby limiting the effectiveness of debiasing. To this end, we propose debiasing service recommendation via causal embedding learning (Descale), which addresses popularity bias by modeling the joint influence of global popularity and personal popularity within a counterfactual inference framework. First, we utilize large language models to reason over and summarize textual information, uncovering complex semantics of user behaviors and preferences, and integrate it with latent representations learned from user-service interactions to construct more comprehensive user representations. Next, we define personal popularity based on similar user groups, where each similar user is assigned a differentiated weight according to its similarity to the target user. This mechanism distinguishes diverse service demands across users and prevents highly active users from dominating the recommendation results. Finally, through counterfactual inference, the influences of global service popularity (GP) and personalized popularity (PP) are jointly calibrated. Empirical evaluation on three real-world datasets demonstrates the effectiveness of Descale in debiasing service recommendations.
Edge-deployed large language models (LLMs) face a critical trustworthiness challenge: constrained by model capacity, edge LLMs exhibit substantially higher hallucination rates on knowledge-intensive queries than their cloud counterparts, yet existing edge-cloud scheduling methods route requests based solely on latency or load balancing, lacking any awareness of—or guarantees on—semantic correctness. Reliable routing hinges on quantifying the semantic risk of edge outputs and providing statistical guarantees on error rates. Semantic entropy effectively quantifies semantic risk, and online adaptive calibration—inspired by conformal prediction—offers principled error rate control under distribution shift. However, conventional semantic entropy computation requires multiple sampling passes ($\approx$6 s), which is infeasible on edge devices; moreover, static threshold calibration cannot adapt to distribution shift, causing actual error rates to far exceed the target. To overcome these bottlenecks, we propose RACE (Risk-Aware Cloud-Edge Orchestration), the first framework to integrate online adaptive threshold calibration with semantic risk budgeting, providing provable bounded ($T$-independent) regret and finite-sample coverage guarantees. RACE introduces three key innovations: a Semantic Entropy Probe (SEP) that predicts semantic uncertainty from prompt-only hidden states in a single forward pass ($\lt $1 ms), enabling real-time hallucination detection on edge devices; an online adaptive calibrator that fuses risk control with online convex optimization to dynamically adjust decision thresholds under distribution shift; and a semantic risk budget grounded in Lagrangian duality theory that bounds the cumulative hallucination rate below a user-specified tolerance $\alpha$. Experiments on MMLU and TriviaQA with GPT-5 as the cloud worker show that, with Phi-3-medium as the primary edge model, RACE achieves a cloud invocation rate of only 4.05–6.25% while keeping edge error rates below $\alpha {=}30\%$—the lowest invocation rate among the risk-controlled methods (Static-CP, ACI, L-ARC); experiments with the weaker Mistral-7B further expose the feasibility boundary of $\alpha$ and the role of the adaptation budget $\rho$.
IP state traceback is a necessary emergency incident response technology, which can help Internet Service Providers (ISPs) to capture the footprint of any traveled IP flow at the Autonomous System (AS) level and reconstruct inter-AS routing path to locate the source. To do this, it requires each AS to establish the routing state fingerprints for as many flows as possible. However, when taking the autonomy of AS, its limited flow processing capability and routing state privacy into consideration, existing schemes suffer the following drawbacks: lack of economic incentives for adoption, lower traceback accuracy and routing fingerprint leakage. For this, we propose NET-Tracker, a service oriented IP state traceback. In NET-Tracker, we adopt blockchain oracle technique to build an inter-AS traceback service trading platform, which can allows the requester to automatically and securely complete fingerprint transaction with participating ASes. And this well addresses the economic issue for its practical deployment. Moreover, with the restriction of fingerprint generation capacity on routers, we devise an AS-level fingerprint establishment optimization algorithm to maximize the fingerprints established in each AS, and thus improve the overall traceback accuracy. Furthermore, considering that the attackers may exploit the blockchain feature of open and transparent transaction to initiate variety of strategic attacks, we devise a privacy-preserving protocol for NET-Tracker. We have conducted the rigorous theoretical analysis and the extensive simulation experiments based on real-world topology to evaluate the performance. The results show that its traceback accuracy could be well anchored at levels close to 97%.
Fine-tuning large language models (LLMs) on private datasets is an effective approach for building domain-specific services, but outsourcing training may expose sensitive data, intermediate activations, gradients, and trainable parameters to untrusted infrastructure. Blockchain can provide transparent governance, while Trusted Execution Environments (TEEs) offer attestable execution; however, TEE CPU throughput alone cannot sustain LLM-scale matrix computation, and direct GPU offloading exposes sensitive tensors. We present PriPTune, a privacy-preserving LoRA fine-tuning framework that combines on-chain governance with TEE-GPU collaborative computing under explicit TEE and side-channel assumptions. Smart contracts manage initialization, authorization, evidence anchoring, and model release, while off-chain TEEs enforce confidential execution. PriPTune keeps raw activations, gradients, LoRA adapters, optimizer states, and mixing secrets inside the TEE, and offloads only frozen-weight GEMMs over block-mixed tensors to an untrusted GPU. A TEE-held permutation and block-mixing transform hide raw token-level activations and avoid the same-order TEE-side correction required by additive masking. PriPTune further supports probabilistic GPU-result auditing to detect manipulation with configurable overhead. Evaluations show that PriPTune recovers ROUGE-L and BERTScore within 1.6% of plaintext LoRA, reduces protected decoder-block latency from a CPU-only upper bound of 1091.8 ms to 30.4 ms, and provides practical blockchain confirmation and throughput for lifecycle governance.
The rapid expansion of the cloud computing ecosystem necessitates scalable and energy-efficient resource management. However, achieving optimal energy efficiency is severely constrained by the complex interplay between computational workloads and thermal dynamics. Traditional data center (DC) scheduling methods typically rely on computationally expensive mechanistic models, such as computational fluid dynamics (CFD), or inefficient algorithm search, which frequently yield suboptimal solutions and struggle to adapt to real-time environments. To address these challenges, we propose an ecology-conservation-optimization virtual machine scheduling (EcoVMS) framework, a data-driven and end-to-end framework for compute-energy co-optimization. At its core, EcoVMS utilizes a latent reward generation enhanced proximal policy optimization (LARGE-PPO) algorithm, which utilizes large language models (LLMs) to analyze runtime telemetry-including central processing unit (CPU) utilization, thermal boundaries, and power consumption, and then utilizes large-scale semantic priors to generate potential reward factors. This mechanism resolves the sparse reward and credit assignment problems inherent in deep reinforcement learning without requiring complex physical modeling. Extensive simulations demonstrate that EcoVMS significantly outperforms baseline methods across diverse workload scenarios. Under a typical 75% workload, EcoVMS achieved a 12.4% improvement in significant energy savings, improving computing resource utilization while reducing hotspots and lowering service latency, resulting in guaranteed quality of service (QoS).
Knowledge distillation (KD) is a widely utilized technique for transferring knowledge from a large model (the teacher) to a smaller model (the student). Owing to its flexibility and broad applicability, KD has been extensively applied in the compression of server-side models to meet the Quality of Service (QoS) requirements of client users. Despite significant advancements, the performance of distillation is substantially compromised when a large disparity exists between the capabilities of the server and the requirements of the client. To alleviate this problem, we propose a novel distillation approach, named Progressive^2, which operates through the combination of a progressively stronger teacher and a progressively smaller student. On the side of the teacher, rather than involving all layers simultaneously, we progressively select additional layers for distillation following a raw-to-rich semantic progression, establishing a systematic learning curriculum. Furthermore, we design a teacher-side multi-feature fusion adapter for the teacher to improve training stability, which is theoretically supported by the framework of Lipschitz continuity. On the side of the student, rather than directly training a tiny model, we gradually reduce the size of the network to facilitate an iterative co-evolution with the teacher. Progressive^2 serves as a flexible framework; the progressive strategy of the teacher can be deployed independently to achieve an optimal balance between accuracy and training efficiency, while the joint integration of the teacher and the student yields further improvements in overall performance.
Anonymous communication services represent a critical class of privacy-preserving infrastructure within the service computing paradigm, where trustworthiness, resilience, and quantifiable quality-of-service (QoS) are foundational requirements. Tor-the most widely deployed anonymous communication service-relies on a centralized Directory Authority (DA) component that functions as a monolithic trust anchor. This architectural choice introduces severe service-level vulnerabilities: single points of failure degrade availability (QoS), metadata exposure compromises privacy (a core service attribute), and centralized control enables censorship-directly violating service computing principles of decentralization, fault tolerance, and user-centric trust. We propose AuDi-BCoRN (Automatically Distributed Blind Choosing of Relay Nodes), a service-native re-architecture that transforms Tor's directory layer into a decentralized, self-governing service ecosystem grounded in the tenets of service computing. By integrating blockchain as a decentralized trust service and using smart contracts for automated service coordination, AuDi-BCoRN eliminates centralized dependencies while closely matching native Tor performance: on a 3% tornettools-generated Tor topology, the empirical CDFs of TTFB, latency, and throughput are nearly indistinguishable from Tor beyond the median, while demonstrably strengthening service integrity against DoS, censorship, and traffic-analysis threats.
The inherent complexity of microservices makes them prone to cascading failures, yet locating the precise root cause remains a critical challenge. Existing multimodal methods often fuse logs, metrics, and traces through coarse feature concatenation, which can introduce semantic interference. More over, although some recent studies have advanced from service level localization to metric-level localization, they usually rely on static correlation or causal models and thus struggle to capture dynamic fault propagation. We present MicroInspect, a progressive framework for metric-level root cause localization in microservices. MicroInspect first disentangles multimodal features to reduce semantic interference and dynamically fuses them with a Mixture-of-Experts gate adapted to fault propagation stages, supporting accurate anomaly detection and service-level localization. It then employs a Dynamic Neural Granger Causal ity Network (DNGCN), guided by anomaly detection signals, to infer time-varying metric dependencies and progressively refine root causes from faulty services to metric-level causes. Evalu ation on four public benchmarks (TrainTicket, SocialNetwork, FAMOS-Mall, and GAIA) shows that MicroInspect achieves the best overall performance across anomaly detection, service level localization, and metric-level root cause localization, with statistically significant improvements in F1-score and HR@1 under paired t-tests (p < 0.05).
Mobile edge computing (MEC) has accelerated the development of artificial intelligence and Internet of Things technologies, leading to the explosive growth of intelligent applications characterized by resource intensity and latency sensitivity, such as image processing and smart home. In practice, an application typically consists of multiple tasks with execution dependencies, where the output of some tasks serves as the input for specific others. Recently, the design of computation offloading methods for such execution-dependent tasks has received extensive research. However, computation offloading for execution-dependent tasks with service dependencies in resource-constrained multi-user, multi-edge-server cooperative MEC systems has not been thoroughly studied. In this paper, we formulate a cooperative computation offloading problem for dual-dependency tasks in multi-edge-server scenarios with limited service and computing resources, aiming to minimize the long-term average service delay for multiple users. To solve this problem, we propose a recurrent multi-agent reinforcement learning-based dual-dependency task offloading (RMA-DepO) algorithm, which enables users to communicate during training to explore and learn optimal joint task offloading and computing resource allocation strategies, and to make distributed offloading decisions at execution time. Simulation results demonstrate that the proposed RMA-DepO algorithm outperforms several baselines under different network settings, demonstrating its effectiveness in coordinating edge resources for cooperative computation of dual-dependency tasks.
Complex microservices topology poses significant challenges to resource scheduling. Consequently, most schedulers rely on extensive offline analysis to prevent potential SLO violations resulting from inaccurate online decisions, leading to high deployment costs. Furthermore, regarding metric selection that is important to bridge the scheduling decision to actual resource allocation, although CPU throttling-based auto-scaling has emerged as a promising approach to capture latency sensitivity, existing methods lack a quantitative mechanism to map throttling signals to resource quota, relying instead on coarse heuristics that potentially waste resources. To address these issues, we propose Bayesian Optimization and Throttling based Vertical Pod Autoscaling (BOTVPA), a sample-efficient scheduling method. BOTVPA features a two-tier architecture: 1) Central Scheduler: We employ Bayesian Optimization (BO) to enable rapid decision-making with minimal samples, eliminating offline training overhead. 2) Local Scheduler: We introduce a novel CPU throttling model that mathematically bridges the gap between throttling rates and CPU allocation. Leveraging this model, the local scheduler dynamically translates BO's decision into precise CPU resource limits at runtime. Results show that BOTVPA reduces CPU consumption by up to 19.4% over the best-performing baseline while satisfying SLO.
Web API recommendation is crucial for mashup development, yet the rapid growth of service repositories causes severe information overload. Existing methods heavily exploit historical API co-occurrences, which can introduce spurious correlations bias across heterogeneous contexts and aggravate sparsity by ignoring valid but unobserved combinations. We pro pose CICA, a causality-inspired, context-adjusted graph learning framework that (i) disentangles functional semantics from con textual confounders and employs a Query-Aware Graph Neural Network to selectively activate context-relevant edges for dynamic noise filtering. Furthermore, the framework (ii) enriches sparse nodes with implicit semantic similarity edges to alleviate the spar sity bias. Finally, a causal adjustment mechanism is implemented to integrate functional matching with co-occurrence patterns by dynamically controlling the intervention of co-occurrence bias. Comprehensive experiments conducted on the real-world ProgrammableWeb dataset demonstrate that CICA consistently outperforms state-of-the-art baselines across standard evaluation metrics, including Precision, Recall, and NDCG.
The rapid growth of Wireless Body Area Network (WBAN) and the increasing demand for real-time healthcare services have emerged fog computing as a promising paradigm to reduce latency and energy consumption in health monitoring. However, the growing need for computation-intensive services increases computational load on Fog Servers (FSs) and WBANs, leading to high latency and energy costs. Nevertheless, existing approaches often optimize either latency or energy while neglecting Healthcare Service Providers' (HSPs) profit and the diverse computational needs of patients. Thus, this paper presents a fog computing-enabled WBAN framework that minimizes patients' latency and energy costs while maximizing HSPs' profit by formulating utility maximization as an NP-hard problem, where utility is a weighted combination of latency, energy costs, and profit, while prioritizing critical patients. To solve this, we design a matching and exchange-based suboptimal algorithm with polynomial-time complexity O(PF logPF,P2), where P and F denote the number of patients and FSs, respectively, leveraging the concept of preferences. Furthermore, we introduce a dynamic pricing scheme that allows patients to pay in proportion to their computational needs. Extensive simulations and analyses using real-world health data demonstrate that the proposed approach achieves an average utility of 99.01% of the optimal solution.
In the metaverse, human-agent interactions increasingly take diverse forms, such as natural dialogues, which enhance user experience but pose challenges for precise service delivery, as users often provide incomplete or ambiguous information-a phenomenon we term fuzzy requirements. This paper addresses how to clarify fuzzy requirements through multi-turn interactions. We formally define the task, construct a benchmark across five domains with user profiles, and propose Tower, a user experience-oriented decision-making algorithm for fuzzy requirement clarification. Tower organizes service parameters into a feature tree as a search space for inquiries and refines it using user profiles and diverse interview questioning strategies. To enable realistic evaluation, we develop a user simulator based on action selection, modeling user personality traits using prospect theory and Fitts's law from cognitive psychology. The simulator proactively presents fuzzy requirements and responds adaptively according to personality traits. Using dialogue histories between Tower and the simulator, we train TowerLM, an automatic clarification model with strong domain generalizatio. Experimental results show that both Tower and TowerLM achieve state-of-the-art performance across three user personality types and can be seamlessly integrated as plugins into existing service delivery methods, improving the delivery success rate under fuzzy requirements on four public datasets.
With the rapid development of 6 G communication technology and the Internet of Things (IoT), mobile edge computing (MEC) is regarded as an effective paradigm of providing low-delay, high-quality services to mobile users. In the IoT device-edge-cloud network, the optimal deployment of MEC servers is a prerequisite for a better task offloading, while the improved performance of mobile users task offloading also indicates the deployment scheme is optimal. Most of current MEC servers deployment studies focus on reducing delay and deployment costs, but ignore the offloading requirements of mobile users with similar task type and cooperative relationship arriving at the same community. In this paper, we study the MEC servers deployment driven by the task offloading requirements of community mobile users in current period by utilizing the stability of their social cooperative relationships to maximize the service satisfaction of all community mobile users in the future task offloading. First, the cooperative relationship strength between mobile users is measured to form a group of resource requesters based on interaction probability, movement trajectory and credit strength. Then, we implement the optimal search of base stations (BSs) using spatial index, followed by the one-to-many matching theory between BSs and community group resource requesters, to balance the load of BSs and reduce the communication delay between them. Finally, we use TD(lambda) algorithm and task similarity between cooperative users to deploy MEC servers with suitable resources around BSs so that the deployment scheme can significantly improve the future task offloading performance of all community mobile users. Based on the real data set provided by Shanghai Telecom, it is confirmed that the proposed scheme has significant advantages in improving all community mobile users service satisfaction, with an average improvement of 18.49% compared with the baselines.
Partitioning algorithms are crucial in distributed stream computing, as they directly affect load balancing across downstream task instances, cluster performance utilization, and data stream processing efficiency. However, existing algorithms suffer from key limitations: they fail to effectively balance load imbalance and key-splitting overhead as excessive splitting increases aggregation costs and memory usage, while insufficient splitting fails to reduce skew. In addition, they cannot timely identify newly emerging high-frequency keys, because historical frequency records dominate detection, which worsens load im balance over time. To address these limitations, we propose Sa Stream, a skewness-aware partitioning framework that rapidly captures changes in stream skewness and dynamically adjusts partitioning decisions, thereby balancing load imbalance and key splitting overhead under evolving stream distributions. (1) A key frequency decay mechanism progressively reduces the influence of historical frequency statistics to improve the prediction of evolving key frequencies. (2) An adaptive threshold modulation method dynamically adjusts the hot-key threshold based on the current key frequency distribution, balancing load imbalance and aggregation costs. (3) A multi-tiered key segregation mechanism classifies keys into cold, warm, and hot categories: cold keys use deterministic hash routing to minimize aggregation costs, hot keys use degree-adjusted key splitting to reduce load imbalance, and warm keys use one-time key splitting to bridge the two extremes. Experiments conducted on Apache Storm demonstrate that Sa Stream outperforms existing stream partitioning approaches, increasing throughput by up to 16% and reducing latency by up to 50%. Additionally, key splitting is reduced by up to 18%, and load imbalance is decreased by up to 98%.
Foundation model inference is rapidly turning into a pay-per-request utility, yet today's infrastructure still deploys experts as dedicated, always-on services. In Mixture-of-Experts models, sparse activation makes per-request execution paths highly dynamic, which amplifies provisioning uncertainty while creating opportunities for Inner-Service Resource Sharing (ISRS) across requests. This paper studies cost-efficient provisioning of MoE inference as a service over geo-distributed edge infrastructure, where a provider must satisfy the Service Level Agreement (SLA) constraints while minimizing the joint cost of compute, VRAM, and bandwidth. We formalize the Mixture-of-Experts Deployment (MoED) problem by explicitly modeling ISRS under dynamic routing and multicast communication. To solve MoED, we design an Expert Cost Loss (ECL) metric and a Sparsity Aware Augmented Graph (SAG) that together guide ISRS-aware expert placement and bandwidth reservation, and we propose the ECL-based Expert Placement in SAG (ECL-EPS) algorithm. We prove that ECL-EPS achieves a logarithmic approximation ratio under practical assumptions. Extensive simulations show that ECL-EPS consistently reduces deployment cost and improves resource-sharing efficiency compared with the benchmarks.
Workflow scheduling in heterogeneous cloud environments must operate under dynamic and heterogeneous conditions. Heuristic algorithms cannot adapt online to changing conditions, while single-agent deep reinforcement learning (DRL) faces issues such as exponential growth in the action space, excessive training time, and limited interpretability. This paper proposes MAR-XGB, a hybrid scheduling framework that combines multi-agent reinforcement learning (MARL) with an XGBoost predictor. The tree ensemble estimates task runtime and execution cost, from which an urgency score is derived. Each agent uses lightweight Monte Carlo Tree Search (MCTS) to select VM locations and make online placement decisions over the Top-k tasks, ensuring real-time adaptability while reducing action-space complexity and improving prediction transparency and decision auditability. Experiments on representative WfCommons scientific workflows and controlled high-load DAG scenarios show that our approach outperforms traditional and state-of-the-art schedulers in terms of makespan, resource utilization, and cost efficiency. As the complexity and scale of workflows increase, our approach demonstrates improved scalability, indicating its potential to handle dynamic cloud workloads with fluctuating demand.
