Driven by the growing demand for computational resources, the computing power network (CPN) model has emerged to enable more flexible and collaborative resource sharing. In this work, we envision a federation of independent computing power providers (CPPs) that collaborate to monetize their idle capacity. A computing power broker is introduced as an intermediary to coordinate resource transactions within the federation. However, the interplay of dynamic pricing between providers, fluctuating demand, and resource heterogeneity poses significant challenges to request scheduling to achieve long-term profitability and market stability. To address this, we propose GRASP, a three-tier framework for adaptive resource reservation and request scheduling in a computing power federation. GRASP employs Proximal Policy Optimization (PPO) to learn global reservation strategies, in which a virtual-queue-based reward shaping term inspired by Lyapunov optimization discourages persistent service deficits. A greedy strategy determines the reservation quantities per-provider, followed by an optimization based method that allocates resources to requests. Extensive experiments show that GRASP consistently achieves higher long term profit than both classical allocation heuristics and stronger online reservation baselines. Sensitivity analyses on pricing and supply, together with a utilization decomposition and high-value service-quality metrics, further explain where this advantage comes from.
Large language model (LLM) serving systems must handle highly fluctuating workloads under limited GPU resources. Fixed provisioning either causes SLO violations during bursts or wastes resources during low-load periods, while dynamic GPU scaling is often too slow to react. Model quantization provides a knob to adapt resource demand to workload dynamics, but existing load-aware quantization techniques overlook the heterogeneous impact of quantization across user tiers and workload types, incurring excessive quality degradation and reducing the perceived value of premium service tiers. Workload Fixed GPU Capacity GPU Scaling Overload Period Workload Fluctuating Workloads Overload Periods rise delay fall delay Fixed GPU Capacity GPU Scaling Time Dynamic Precision We present DPServe, a dynamic-precision LLM serving frame work for multi-tier SLO preservation under fluctuating work loads. DPServe jointly optimizes all users' SLO satisfaction and free-user quality preservation by selectively adjusting the precision of free-user requests across requests, generation stages, and task types. To avoid excessive degradation, DPServe introduces a quantization budget that mitigates long-term response quality loss and compensates repeatedly degraded users with higher precision when resources permit. By coordinating quality-aware scheduling with a hybrid multi-precision LLM instance, DPServe adapts resource demand to workload changes without dynamic GPU scaling, preserving paid-user SLOs while protecting free users from unnecessary response quality loss. We evaluate DPServe across four representative LLMs. DPServe reduces P95 TTFT by 2.33×–7.41× compared with the full-precision and DVFS-based baselines while limiting task-quality degradation to 0.7%–2.0%. Compared with MorphServe, DPServe reduces response quality loss by 1.96×–3.30×, with an increase of only 0.613–2.588 s in P95 TTFT.
Recently, the rapid increase of maritime activities brings significant challenges such as limited computing resources of maritime devices (MDs) and unreliable maritime communication links. To address these issues, this paper proposes a simultaneously transmitting and reflecting reconfigurable intelligent surface (STAR-RIS)-enhanced unmanned aerial vehicle (UAV) assisted maritime mobile edge computing (MEC) system, where a STAR-RIS is mounted on UAV to transmit uplink signals from MDs to UAV and reflect them to a coastal base station (BS), thus UAV and BS can cooperatively process the offloaded tasks and allocate appropriate computing resources. Besides, we define a System Overhead Ratio (SOR) metric to characterize the trade-off between system latency and energy consumption, and formulate a SOR minimization problem by jointly optimizing partial offloading, STAR-RIS phase-shift, resource allocation, and UAV trajectory. Considering the non-convex and hybrid discrete-continuous nature of the problem, we propose a Stability Enhanced Hybrid Proximal Policy Optimization (SE-HPPO) algorithm, which effectively stabilizes the policy update process in a high-dimensional hybrid action space by incorporating dual KL divergence constraints and an adaptive clipping mechanism. Numerous simulations demonstrate that the proposed SE-HPPO algorithm outperforms baseline schemes in terms of convergence and SOR under various system configurations. These findings support the feasibility of leveraging UAV-mounted STAR-RIS to improve maritime communication environment and empower aerial-ground cooperative MEC paradigm.
Computing power networks integrate multi-supplier heterogeneous resources to support diverse computational tasks. However, the following issues persist: elevated management overhead due to divergent object description standards among suppliers; inefficient retrieval caused by the use of traversal-based methods for massive resources; and inadequate balancing of multi-stakeholder interests stemming from variations in the quantity and nature of their optimization objectives. To standardize object descriptions, a unified characterization model and a three-tier service architecture are proposed. Additionally, a dual-feature-aware resource allocation approach utilizing learnable tree navigation is presented. This approach employs a learnable resource directory tree enhanced by a reinforcement learning agent for rapid resource retrieval, and determines the optimal allocation scheme through game-theoretic interaction between resource providers and users to balance their conflicting interests. The feasibility of this approach has been rigorously demonstrated through theoretical derivation. Experimental results confirm the efficiency of the proposed methods and system: the retrieval method reduces resource retrieval time by over 90% across all characterization scenarios; the Characterization Manager cuts processing time by 80% across three core functions; and the scheduling method improves the total interest score by 10–35% across standard workflow tasks of varying scales, outperforming the best benchmark in interest balance.
Programmable data planes enable in-network computing functions in cloud fabrics, but their prevailing compileand- redeploy workflow creates a reconfiguration bottleneck: application logic remains coupled to a target-specific pipeline layout, updates can require disruptive base-pipeline redeployment, and switch memory is reserved statically at compile time. We propose P4asm, an eBPF-inspired runtime-virtualization framework for supported in-network computations. P4asm installs a fixedparser P4 runtime that interprets compact virtual-instructionset bytecode and manages application code and state in a logical heap. The controller can then install, replace, and reclaim supported application logic without redeploying that fixed base runtime. We implement the framework on the BMv2 behavioral model to validate its functional design. Our BMv2 testbed results provide prototype-level evidence rather than ASIC claims. In representative traces and request replays, P4asm reduces application code size by 13× to 25×, avoids reboot outages during logic injection, keeps the critical version switch step within 10 ms to 20 ms across three sampled update cases, and preserves bounded behavior under loss and controller disconnects. The workload and ablation figures are reported as fixed operatingpoint summaries of these testbed runs.
Long-term cloud workload forecasting is crucial for proactive resource scheduling and efficient cloud service management. Traditional forecasting methods, including recent Transformer-based approaches, often struggle to capture the intricate patterns and dynamic distribution shifts inherent in long-term workloads. To address these challenges, we propose CoTF, a robust representation learning framework for long-term workload forecasting. Our approach begins by treating historical and future workload segments as distinct views and leveraging contrastive learning to derive discriminative representations that mitigate distribution shifts. Next, a self-adaptive decomposition strategy disentangles complex workload patterns by extracting temporal and frequency domain features from each view. Finally, a fusion module integrates these temporal-frequency representations via bidirectional cross-attention, resulting in more accurate predictions. Extensive experiments on real-world cloud workload traces demonstrate that CoTF significantly outperforms state-of-the-art methods, achieving average improvements of 15.02% in MAPE and 13.23% in RMSE. The code is available: https://github.com/QiuyueLv/CoTF.
Federated learning protects user privacy by sharing only local model parameters rather than raw data, but it is vulnerable to poisoning attacks. Existing defenses rely on plaintext gradient analysis, posing privacy leakage risks; while encryption, though secure, incurs high computational overhead, making detection unaffordable. To address these issues, this paper proposes a Privacy-Preserving and verifiable federated learning scheme with lossless weighted aggregation (PPVFL-LWA), which effectively resists poisoning attacks and supports user dropout while ensuring user privacy. The scheme protects local gradients with obfuscation factors to prevent collusion between the server and users for privacy inference; ensures data integrity and lossless aggregation of local models through a dual-collector mechanism and bilinear aggregation signatures; and introduces verification contracts to eliminate reliance on trusted third parties for verification. Furthermore, this paper designs a semi-trusted dynamic weight adjustment strategy based on gradient similarity and data distribution characteristics to mitigate the impact of poisoning attacks by adaptively adjusting participant weights. Experimental results show that PPVFL-LWA outperforms existing schemes in both training efficiency and attack resistance.
Existing switch-based load balance schemes in data center networks face a fundamental dilemma: local strategies respond quickly but cannot detect downstream congestion, while global strategies approximate network-wide views but suffer from information propagation delays. These issues prevent data flows from quickly and accurately bypassing congested links, resulting in prolonged Flow Completion Time (FCT). We propose Label Based Dynamic Load Balance (LDLB), a distributed algorithm that bridges this gap by combining instantaneous local congestion detection with globally synchronized congestion labels. LDLB uses Random Early Detection and Explicit Congestion Notification (RED/ECN) to identify congested egress ports and records them in a local congestion table (LCT). Congested switches proactively generate congestion labels—identifying affected destination IPs—and forward them to upstream neighbors, which cache the labels in a congestion label table (CLT). A compressible information table aggregates contiguous IP addresses to reduce storage and synchronization overhead. Upon packet arrival, each switch consults both the LCT and CLT and applies a Weighted Cost Multi-Path (WCMP)-like cost function to reroute traffic away from congested links. Experiments on ns-3 with four real data center workloads (Web Server, Cache Follower, Web Search, Data Mining) demonstrate that LDLB reduces FCT by 57%, 61%, 65%, 67%, and 54% compared with ECMP, DRILL, LetFlow, HULA, and CONGA, respectively.
Blockchain, originally introduced for cryptocurrencies, provides a secure and decentralized platform for processing and storing transactions among distributed cloud nodes. However, traditional blockchain systems require broadcasting each transaction to all nodes, incurring high computational and communication overhead. In cloud networks, this problem is exacerbated because forwarding transactions consumes limited bandwidth and computing resources, which can discourage nodes from participating honestly. In addition, some nodes may refuse to forward transactions, undermining network consensus. In this paper, we propose Incentive Transaction Forwarding in Cloud networks (ITFC), which rewards cloud nodes with a share of transaction fees for forwarding transactions. We design a dynamic topology update mechanism to enable accurate incentive computation and develop a specified blockchain structure to support fair revenue distribution. We analyze ITFC under a stated rational-adversary model, including Sybil and active-user-node strategy, and establish a local monotonicity property for link removal that preserves the shortest-path level assignment. Extensive simulations evaluate topology-based incentive allocation, the two modeled attacks, heterogeneous cloud-node capacity, and algorithmic computation and storage overhead.
Intelligent parking systems significantly improve urban traffic efficiency by reducing parking space search time and increasing resource utilization. However, the continuous leakage of parking data has brought concerns to people. To address this issue, this article proposes a secure and reliable parking data outsourcing range query privacy preserving scheme for the Internet of Vehicles. This scheme is based on the alliance chain as the information exchange platform and provides outsourced query services through smart contracts. Taking into account the personalized query needs, the spatiotemporal aggregation scheme is deeply designed and implemented. Through time-based aggregation, the addition or multiplication of ciphertext data in different time periods can be better realized. Through space based aggregation, the ciphertext data calculation of different users or the same user in different parking lots can be better realized. In order to improve query flexibility, a hybrid encrypted query algorithm is designed, which can not only provide encrypted numerical range queries, but also encrypted text matching range queries. In addition, reliable incentive mechanisms are implemented and personalized incentive measures can be implemented. Last but not least, theoretical analysis and simulation experiments are implemented. The results show that the proposed solution can well meet the design requirements of security, privacy, flexibility and reliability.
The rapid growth of cloud computing demand has fueled the expansion of cloud data centers (CDCs). To improve resource utilization and operational efficiency, effective coordination of geographically distributed CDCs is becoming increasingly important. This paper proposes a consensus-based asynchronous decentralized optimization method for geo-distributed workload scheduling across CDCs. Specifically, computing requests are categorized into three classes according to their latency tolerance, based on which a decentralized scheduling architecture is developed. Then, the geo-distributed workload scheduling problem is formulated as an integer linear program (ILP), and further transformed into a decentralized optimization model by introducing consensus variables to represent inter-regional workload transfers. To address heterogeneous communication delays and straggler effects, an asynchronous consensus algorithm is developed. Two novel mechanisms, known as the minimum dependency set and maximum lagging level, are leveraged to balance update responsiveness and information freshness. Numerical studies on exemplary cloud computing systems demonstrate that the proposed scheduling method reduces overall cost by 15.5% compared with the local-only computing paradigm. Besides, our asynchronous decentralized algorithm has reduced the wall-clock time by 53.6% and 47.8% under fixed- and stochastic-delay straggler scenarios, while achieving solution quality close to that of the centralized benchmark.
In the mobile edge computing (MEC)-enabled Internet of Vehicle (IoV) networks, the communication quality between vehicles and edge servers during task computation offloading is vulnerable to degradation caused by obstacle interference, thereby reducing user quality of service. The Intelligent Reflecting Surface (IRS) is a passive and programmable device, can be installed on the exterior of buildings to reflect and enhance signals through phase shift adjustment. However, a coupling relationship exists between IRS phase control and task offloading decisions. Optimizing them separately by decoupling fails to reflect the highly coupled and dynamic interactions in real systems. This may lead to a mismatch between offloading decisions and channel conditions, making it difficult to achieve optimal system performance in real-world environments. To address this challenge, we propose a joint optimization method for IRS phase control and task computation offloading based on causal inference (OITC). Specifically, we first analyze and model the dynamic channel between vehicles and the IRS. Then, causal inference is used to verify the causal relationship between IRS phase control and task computation offloading. Moreover, we combine causal inference with the Multi-agent Proximal Policy Optimization (MAPPO) algorithm to jointly optimize IRS phase control and task offloading decisions. Ultimately, extensive experiments demonstrate that the proposed algorithm outperforms traditional multi-agent deep reinforcement learning (MADRL) methods in terms of communication quality, system delay, and energy consumption.
Cloud platforms offer innovative solutions for data processing, and privacy concerns arising from this can be addressed through secure multi-party computation (MPC). MPC in the random access machine (RAM) model leverages Oblivious RAM (ORAM) to conceal memory access patterns, rendering it more suitable for data storage and aggregation in cloud environments, where cloud servers can infer user privacy based on observed access patterns. However, most existing RAM-based MPC protocols only ensure semi-honest security; the few actively secure alternatives have limitations in efficiency or input leakage. To address this gap, this paper proposes an efficient active-secure RAM-based MPC protocol without input leakage. Specifically, we first present a new actively secure, reactive MPC protocol with communication of 2(n−1) bits per gate. As the building blocks, we design the (i) multi-party distributed BitGC garbling scheme and (ii) multi-party somewhat homomorphic encryption (SWHE) scheme with distributed decryption. In particular, we then build a RAM-based MPC protocol on top of the above new building blocks with Circuit ORAM and present the check mechanism to achieve secure memory accesses. The implementation shows that our protocol outperforms the prior actively secure RAM-based MPC without leakage by one to two orders of magnitude in both LAN and WAN networks. Compared with the state-of-the-art actively secure RAM-based 2PC with one bit leakage, ours is slightly less efficient in most cases (not all), but supports the multi-party setting and offers a superior alternative for highly confidential environments.
The paradigm shift of remote sensing (RS) analysis toward cloud-based large-scale computing demands intelligent methods capable of interpreting petabyte-scale, high-resolution heterogeneous imagery. Multimodal Large Language Models (MLLMs) have emerged as a transformative solution for such data-intensive tasks, offering unprecedented capabilities in open-set perception and high-level semantic reasoning. However, directly applying general MLLM architectures to the RS domain faces significant challenges: the constraint of fixed-resolution vision encoders causes severe information loss for small objects, while the reliance on single-layer feature fusion neglects critical spatial cues preserved in intermediate layers. To address these challenges, we propose ROARS: a Cloud-Based Resolution-Native Framework for Large-Scale Remote Sensing Multimodal Analysis, tailored for high-fidelity interpretation. First, we introduce a Native-Resolution Vision Encoder tailored to process images of arbitrary aspect ratios and scales without downsampling, thereby strictly preserving pixel-level details. Second, we design a Shortcut-Based Fusion Mechanism that bridges the semantic gap by injecting multi-level visual features into the language model. Furthermore, to enhance global contextual understanding, we incorporate a Multi-Granularity Token Pooling strategy that provides fine- and coarse-grained visual representations. Extensive experiments on the VRSBench demonstrate that our framework significantly outperforms state-of-the-art approaches, particularly in fine-grained object perception and complex spatial reasoning. These results validate the proposed method as a robust and efficient core component for next-generation cloud-based RS interpretation systems.
Unmanned aerial vehicle (UAV)-assisted Internet of Vehicles (IoV) improves the robustness of vehicle-to-Road Side Unit (RSU) connectivity in settings with sparse roadside infrastructure and severe urban blockage. However, introducing airborne relays and open wireless links also substantially enlarges the attack surface. Existing authentication schemes for this scenario typically provide only session-level anonymity, while revocation is either omitted or delegated to a single trusted authority, making the system vulnerable to well-known attacks and yielding revocation views that are neither scalable nor consistent across RSUs. To address these issues, we propose BRAV (Blockchain-enabled Revocable Authentication for UAV-assisted Internet of Vehicles). BRAV enhances self-updating hash chains to enable cross-session pseudonym evolution with public-transcript unlinkability against external observers. Moreover, BRAV introduces a Dual Cuckoo Filter revocation mechanism that leverages the globally confirmed state of a consortium blockchain to support fast, compact, and verifiable revocation queries, while maintaining a consistent revocation view among RSUs. We prove session-key security of BRAV in the Real-or-Random model and validate its authentication and secrecy properties using Scyther. Experimental results show that BRAV maintains practical computation and communication overhead while providing stronger security coverage than lightweight baselines. Its filter-based revocation significantly improves query efficiency over conventional certificate revocation lists (CRLs) and a single-filter baseline, and reduces storage consumption by about 87% compared to CRLs. Finally, NS-3 and blockchain simulations further demonstrate the applicability of BRAV in UAV-assisted IoV deployments.
Data across private clients in edge networks through federated learning (FL) provide crucial resources for the advancement of artificial intelligence (AI), but are often in a non independent identically distributed (non-IID) manner. As FL clients only provide local models to the task publisher, their information asymmetry, such as training data sizes and non IID degrees, significantly impacts on the global model accuracy. Therefore, this paper address the problems of how to make use of the clients' training data attributes, select which clients for participating in FL, and devise reward strategies to meet economic conditions and boost engagement of high-quality clients. In particular, we propose a utility-estimation and quality evaluation (Uteque) incentive mechanism for FL. Uteque offers FL efficacy through three components: 1) data classification and publication for revealing true training data information of clients, 2) utility estimation and client selection for determining the set of clients to participate in FL, and 3) reward determination and learning quality evaluation for paying clients with truthfulness and individual rationality. This way, it evaluates the learning quality of clients solely based on their local models and formulates reward strategies based on auction theory. Extensive experiments on real-world datasets verify the effectiveness of Uteque for FL
Cloud storage requires fine-grained access control (FGAC), efficient permission revocation, and security against quantum attacks. Existing puncturable attribute-based encryption (PABE) schemes either rely on pairing-based assumptions or suffer from key expansion, ciphertext expansion, or owner coupled key management when deployed in multi-tenant cloud environments. This paper presents LPABE, a lattice-based PABE scheme that combines policy-level attribute authorization with tag-level user-autonomous puncturing. The proposed scheme introduces a noise-controlled decryption structure based on disjunctive normal form (DNF) policy transformation and symmetric Lagrange coefficient allocation, reducing reconstruction induced noise growth under the considered policy class. It sup ports puncture-history-independent local puncturing, while user keys and ciphertexts remain independent of puncture history and revocation-list size. A decoupled credential design further enables a single user key to be used across authorized data sources without data-owner-specific rebinding. Compared with representative lattice-based PABE schemes, the construction avoids puncture-count-dependent quadratic key growth, revocation-list dependent ciphertext expansion, and data-owner-dependent key management growth. We prove indistinguishability under chosen plaintext attacks (IND-CPA) in the selective access-policy and selective tag setting under the Learning With Errors (LWE) assumption, and analyze resistance to collusion attacks. Experimental results demonstrate that LPABE provides practical online performance, with decryption remaining below 45 ms for 50 satisfied attribute leaves and below 11 ms for 50 ciphertext tags, while local puncturing stays nearly constant at about 13.6 ms across repeated punctures.
