With the rapid advances in Large Language Models (LLMs), LLM inference has emerged as a critical research domain. Edge cloud systems have become the preferred platforms for deploying models due to their proximity to end users. Meanwhile, an LLM inference request typically requires multiple and interdependent models, and is also time-sensitive. Thus, how to ensure the timeliness of inference across interdependent models deployed on distributed edge cloud systems remains highly challenging. We quantify the system timeliness of model inference by a metric Age of Model Inference (AoMI). By thorough investigation, we consider that there are two challenges in optimizing this metric, including a lack of model deployment adaptability and accurate model invocation decisions. In this work, we propose a targeting AoMI optimization for joint model inference in edge cloud systems, namely TurboInfer. To tackle the challenge of lacking deployment adaptability under dynamic, highly coupled request-model relationships, we model the request-model interactions as a hypergraph and apply hypergraph partitioning with a two-stage model redeployment mechanism to balance communication and migration costs. To address the challenge of inaccurate model invocation decisions due to complex model dependencies, we introduce a memory association learning method that enhances sample efficiency. The evaluations using real-world traces demonstrate that our method improves the timeliness of LLM inference.
Advanced Persistent Threats (APTs) span multiple stages, hosts, and time windows, making end-to-end attack reconstruction from isolated audit events challenging. Provenance graph construction models causal dependencies among system entities; however, in distributed environments, naive event linking produces dense graphs, and centralized analysis scales poorly over large audit logs. This paper presents a multi-LLM framework for provenance graph construction over distributed audit logs. It combines role-specialized agents with Tool-Mediated Context Reduction (TMCR), enabling each agent to access only task-relevant evidence instead of large raw log streams. The framework further incorporates semantic-aware edge selection and anchor-based global correlation to suppress redundant relations and reconstruct cross-host attack paths. Experiments on the APT29 audit-log scenario show that the proposed framework improves edge precision and attack-path reconstruction while significantly reducing graph complexity and token usage compared to single-LLM and multi-LLM baselines without TMCR.
The quantum Internet aims to distribute entanglement among geographically separated quantum nodes, enabling applications such as quantum key distribution, distributed quantum computing, and cooperative quantum sensing. Since entanglement generation is probabilistic and quantum memory and concurrency are limited, efficient Entanglement Access Control (EAC) is a key challenge. While a purely quantum Medium Access Control (MAC) scheme using multipartite Dicke states has been proposed to resolve contention without classical round-trip communication, prior work mainly focuses on a single Quantum Local Area Network (QLAN). In wide-area quantum networks composed of multiple interconnected QLANs, supporting hierarchical decision making, dynamic winner counts, and fairness under heterogeneous QLAN sizes remains difficult. In this background, we propose Dynamic Hierarchical Entanglement Access Control (DH-EAC), a two-stage quantum lottery scheme for wide-area networks. In the outer stage, a Global Orchestrator uses a Dicke state to select winning QLANs and determine their quotas.
The rapid deployment of foundation models and AI workloads has significantly increased the energy demand of modern data centers, particularly under edge computing environments characterized by limited resources, heterogeneous devices, and intermittent connectivity. This paper proposes a novel federated digital twin-driven framework for energy-aware scheduling and resilience enhancement in edge data centers. The proposed system integrates digital twin modeling, federated learning, and distributed energy-aware optimization to enable privacy-preserving, scalable, and adaptive model training across multiple edge nodes. To address challenges in communication efficiency, heterogeneity, and system robustness, we introduce a Double Robust Federated Learning (DRFL) architecture that incorporates blockchain-based trust management and distributionaware node selection. Experimental analysis demonstrates that the proposed framework improves energy efficiency, scheduling stability, and system resilience while preserving data privacy. This work highlights a cross-layer design paradigm for federated foundation models operating under resource-constrained edge environments.
Byzantine Fault Tolerant (BFT) protocols are essential for ensuring consistency in decentralized systems. However, once safety is violated, accountability becomes crucial for identifying malicious replicas. Existing accountable BFT protocols often rely on external observers or incur up to $O\left(N^{4}\right)$ communication overhead due to broadcast-based mechanisms. We propose GALUPA, a decentralized accountability mechanism that detects Byzantine misbehavior by leveraging conflicting Quorum Certificates (QCs) exchanged solely among replicas. Compatible with PBFT, HotStuff, and BFT-SMaRt, GALUPA uses a randomized gossip protocol to disseminate aggregated signed locking information. We develop a probabilistic model based on Erdós-Rényi random graphs (ERG) to analyze the accountability success rate under varying network sizes and parameter configurations. Our analysis shows that GALUPA introduces only $O(k N)$ message complexity for any fixed target success rate. We implement GALUPA on the BFT-SMaRt platform and evaluate its performance in networks of up to 160 replicas. Experimental results show that GALUPA maintains over 70% throughput while achieving 99.9999% accountability success, $8 \times$ higher than that of the broadcast-based baseline.
GPU-accelerated quantum simulation is essential for NISQ-era research, but the HPC systems best suited to these workloads expose an operating model oriented toward security and scheduling. That model is poorly aligned with the interactive, device-oriented workflows expected by quantum software. We present HPC-vQPU, a two-plane architecture that exports HPC-hosted simulation as a virtual-QPU service without weakening the site's security posture. A cloud-facing control plane exposes a device-oriented REST API for circuit submission, lifecycle tracking, and event streaming; an unprivileged loginnode agent polls outward, atomically claims tasks, and dispatches scheduler-backed GPU work. A topology- and calibration-aware device snapshot is bound at claim time and carried into each GPU job as a self-contained input, decoupling execution from live server contact while preserving calibration fidelity across scheduling delay. Deployed on Pawsey's Setonix supercomputer with NVIDIA Grace Hopper nodes, Qiskit-Aer/cuQuantum, and real IBM Fez calibration data for a 156-qubit heavy-hex device, HPC-vQPU adds only $2.0-3.4 \mathrm{s}$ of interface-visible overhead in our deployment (3.4-5.9% of end-to-end latency) for 28-32 qubit circuits. The bound snapshot measurably changes output distributions, device state is fixed at claim rather than submission time, and task lifecycle consistency is preserved through agent failure and multi-agent contention without privileged deployment or firewall changes.
Quantum neural networks (QNNs) are an emerging option for autonomous vehicle perception, but their high inference latency makes direct real-time deployment challenging. We present AQUAD, an asynchronous vehicle-to-edge (V2E) framework that integrates a slow quantum perception model into a fast onboard pipeline without blocking local execution. AQUAD combines adaptive offloading with asynchronous quality rescue, allowing delayed edge results to refine ongoing classical perception while preserving responsiveness. We implement AQUAD using PyTorch, TorchQuantum, and multithreaded Python, and demonstrate it on the KITTI dataset through an interactive dashboard showing real-time detections, bird's-eye-view projections, latency, communication overhead, and rescue events. The demo illustrates the feasibility of asynchronous quantum-assisted edge intelligence for real-time autonomous driving.
Distributed computing continuum systems (DCCS) commonly partition application state to enable parallel execution and scale. To avoid costly cross-partition coordination, objects that interact frequently should be assigned to the same partition. Maintaining such efficient partitions under dynamic workloads while keeping updates cheap, however, remains a key challenge. We present SpectralShard, a lightweight partitioning technique based on the spectral structure of the interaction graph. SpectralShard assigns each object a scalar value, derived from a loworder spectral representation of the graph, such that frequently interacting objects receive similar values, and groups objects with nearby values. Because these values can be maintained incrementally as the graph evolves, SpectralShard adapts to workload changes with low overhead, avoiding costly criticalpath computations required by optimization-based approaches. We evaluate SpectralShard on large-scale Ethereum transaction traces as a representative dynamic workload for distributed execution. SpectralShard reduces cross-partition interactions by up to 49.9%, improves throughput by up to 83.5%, and lowers end-to-end latency by up to 38.4% compared to hash-based sharding. These improvements match or exceed those of greedy locality-aware partitioning, while incurring negligible computation on the critical path, comparable to hash-based sharding. Overall, this suggests that low-order spectral methods are a promising building block for adaptive partitioning in DCCS.
Nowadays, uncertainty in data stream analytics for the Industrial Internet of Things (IIoT) makes it difficult to ensure quality of service (QoS). These QoS objectives are usually expressed as constraints on the performance metrics. When runtime contexts fluctuate, and some QoS objectives are no longer met, we aim to achieve steady scheduling, ensuring that other QoS objectives remain met while re-meeting the unmet ones as much as possible. However, searching for a real-time steady scheduling policy in a high-dimensional configuration space is costly, making it difficult to find the best solution in time. To address this issue, we introduce the concept of side-effect, which quantifies the minimal overall QoS loss caused by updating a subset of configurations to re-meet violated constraints. Instead of updating all configurations, we update only a small subset with a small side-effect, thereby shrinking the search space and achieving lightweight, steady scheduling. We have validated the performance of our design in scenarios where both accuracy and delay are subject to certain constraints. Experimental results show that the overall QoS loss achieved by our method remains around 5%, with an average reduction of 46.07% compared to the state-of-the-art (SOTA) methods.
Online data preprocessing has become a major bottleneck for DNN training on resource-constrained edge servers. A promising line of work targets data augmentation, which is often the most time-consuming stage in the preprocessing pipeline. Such approaches alleviate this bottleneck by reusing a subset of augmented samples in future iterations without provisioning additional compute resources. However, existing state-of-the-art reuse methods still incur substantial training inefficiencies, even though they reduce preprocessing time. This paper presents ReCache, a group-based dynamic reuse policy that maximizes CPU-GPU pipeline utilization while preserving model accuracy. ReCache makes three key contributions. First, it derives a reuse threshold that eliminates GPU idle time by balancing CPU preprocessing time with GPU training time. Second, it introduces a multi-dimensional sample assessment mechanism to accurately identify reusable samples. Third, it dynamically partitions training data into groups with fine-grained reuse limits and selectively refreshes stale samples via replacement or reaugmentation to maintain accuracy while minimizing training cost across diverse training settings. We implement and evaluate ReCache across nine models, two tasks on four datasets, and two hardware platforms. ReCache outperforms four state-of-theart methods that do not provision additional compute resources: at the same target accuracy, it reduces training cost by 43% on RTX 4090 and by 30% on V100. Our code is available at https://github.com/optiq-lab/ReCache.
The deployment of directed acyclic graph (DAG)based deep neural networks (DNNs) on mobile devices has become increasingly common, enabling advanced AI applications such as real-time object detection and face recognition. However, running these complex models on resource-constrained mobile hardware presents significant challenges, particularly in simultaneously optimizing inference latency and energy consumption. Moreover, existing frameworks either restrict execution to a single processor, missing out on heterogeneous computing, or lack support for complex DAG-DNNs-failing to meet the needs of modern mobile inference. To address these issues, we propose PipeNN, the first energy-latency co-optimization pipeline for mobile complex DAG-DNN inference. Unlike the conventional paradigm of intra-operator parallelism, PipeNN adopts interoperator pipeline parallelism, enabling efficient task partitioning without extensive profiling and predictor building overhead. To solve the challenging multi-objective optimization of both energy and latency, we propose a coarse-to-fine algorithm that selects optimal partitioning based on user preferences, effectively balancing inference speed and energy consumption. Moreover, we introduce several system-level optimizations to fully unlock PipeNN's performance, including operator building interception, lightweight memory management, and optimized inter-processor pipeline, which significantly reduces storage and synchronization overhead. We implement PipeNN on various platforms and evaluate it on multiple DNN models and devices. Compared to state-of-the-art inference frameworks, PipeNN achieves up to 3.03× speedup and reduces energy consumption by up to 59.7% on complex DAG-DNN inference.
Modern GPU clusters can exhibit underutilization under static allocation and coarse-grained hardware partitioning; in our ResNet experiment, a single workload occupies an entire GPU while utilizing only $65-70 \%$ of compute capacity. Existing elastic scheduling approaches (e.g., AntMan, ElasticDL) rely on Kubernetes-level scheduling and are typically tailored to deep learning workloads, which complicates adoption for heterogeneous applications such as AI4Science, inference serving, and scientific computing. Hardware solutions like NVIDIA MIG provide static partitioning but lack fine-grained elasticity for dynamic workload adaptation. We present FractalGPU, a driver-level elastic virtual GPU scaling framework that enables transparent, fine-grained resource sharing across heterogeneous workloads. FractalGPU implements three tightly integrated mechanisms: (1) an elastic scaling controller that adjusts per-task uplimits online and uses safety bounds to reclaim idle compute capacity without disturbing active execution; (2) a quadratic-feedback delay-based enforcement mechanism that throttles tasks exceeding their uplimits via controlled delays inserted at kernel launch boundaries; and (3) an incentive-aware time-slice scheduling mechanism with rewardpenalty compensation to provide fairness under non-preemptive kernel execution. By operating at the driver layer, FractalGPU requires no application or framework modifications and applies uniformly to CUDA applications spanning DL training, inference serving, and HPC applications. Evaluation on NVIDIA A100 GPUs shows that FractalGPU improves GPU utilization from 65-$\mathbf{7 0 \%}$ to approximately $\mathbf{8 0 \%}$, achieves $\mathbf{2}-\mathbf{7} \times$ higher throughput for LLaMA prompt processing under concurrent workloads, and attains a Jain's index of 0.85-0.95 across heterogeneous task sets.
The security of today's blockchains relies almost entirely on classical cryptography and Proof-of-Work (PoW), both of which are increasingly challenged by advances in quantum computing. Fully replacing PoW with quantum-based consensus, however, remains impractical due to hardware scarcity, high verification cost, and instability under probabilistic validation. This paper presents a hybrid blockchain design that integrates Proof-of-Work and quantum-based consensus into a unified architecture. Our system embeds quantum-validated blocks as periodic checkpoints within a predominantly classical chain, providing quantum-origin security anchors while preserving decentralization, throughput, and miner accessibility. We introduce an epoch-based consensus protocol, a hybrid chainwork formulation that balances classical and quantum contributions, and dynamic adjustment mechanisms that adapt block weights and mining difficulty as quantum hardware evolves. We further evaluate the protocol using discrete-event simulation. The results demonstrate stable protocol operation under probabilistic quantum validation, effective control of the relative contributions of classical and quantum blocks, and predictable incentive behavior across a range of operating conditions. Together, these results demonstrate that hybrid consensus offers a practical and forward-compatible path for evolving classical blockchains toward quantum-resilient operation without disruptive protocol replacement.
With the rapid expansion of link capacity and the increasing performance requirements of emerging applications and services, modern data center networks have widely deployed Remote Direct Memory Access (RDMA) to deliver ultra-low latency and high throughput. However, existing load balancing solutions either suffer from packet loss due to the limitation of hardware resources in RDMA network interface cards, or are far from optimal for mixed-flow parallel transmission. In this paper, we propose a congestion adaptive load balancing with in-network reordering, called CALB, to guarantee that packets arrive at the receiver without reordering, thereby minimizing flow completion time as well as the deadline miss rate. The key idea is that CALB periodically inspects the congestion status of each path while measuring its available bandwidth. Once a flow experiences congestion or one deadline flow may fail to finish transmission within its deadline, the switch assigns one path with available bandwidth matching the congested flow's. Moreover, CALB employs an in-network reordering mechanism to ensure packets arrive at the receiver in order. The evaluation results demonstrate that CALB can reduce deadline miss rate by up to 85% and the flow completion time by up to 65% compared with the state-of-the-art load balancing solutions.
In smart grids, an attacker can reduce his electricity consumption and increase the records of other users with electricity theft attacks. In flat-pricing scenarios, electricity consumption directly decides the final payment. However, in multiple-pricing scenarios, the final payment is determined by multiple factors, which makes the detection of electricity theft more challenging. Existing detection methods are based on strong assumptions that the distributions of users' electricity consumption are independent or highly correlated, which are not suitable for real-world scenarios. In this paper, we introduce a new electricity theft attack, named deceptive electricity theft attack, which allows attackers to be profitable without being detected by existing methods by behaving oppositely to the traditional cases. Specifically, they achieve this by increasing their own meter reading and reducing the meter reading of victims. Additionally, we explore scenarios of deceptive electricity theft attacks involving multiple attackers and victims. To mitigate this attack, we propose an innovative electricity theft detection scheme capable of efficiently identifying both existing and new attacks without relying on strong assumptions, utilizing the bootstrapping approach. Experiments on both synthetic and real-world data show our detection scheme can effectively detect both existing and new deceptive electricity theft attacks.
Large-scale IoT applications increasingly rely on LEO satellite networks to collect geo-distributed sensing data from wide geographic areas. While sketch-based data summarization can significantly reduce storage and query costs, existing approaches typically perform summarization only after raw data are fully collected at centralized data centers, resulting in substantial communication overhead and long end-to-end latency. This paper presents INS, a sketch-based In-Network Summarization framework that jointly schedules data forwarding and compression in LEO satellite constellations. INS enables satellites to incrementally construct sketchlets during data collection by leveraging onboard computing capabilities. To address satellite mobility, heterogeneous resources, and skewed data distributions, we formulate a joint summarization and communication scheduling problem and design a load-balanced compression satellite selection algorithm that minimizes the job completion time (JCT). Simulation results show that INS reduces the JCT by up to 61.6% compared to traditional raw data collection and by 22.2% compared to naïve volume-based in-network compression. The performance gains are particularly pronounced under skewed data distributions and resource-constrained LEO environments.
The proliferation of serverless edge clouds unlocks event-driven function execution across distributed infrastructure. However, heterogeneous resource demands, bursty workloads, and multi-layer orchestration introduce severe provisioning complexity, coordination difficulty, and learning instability. These challenges manifest mainly as resource mismatching caused by coarse-grained readiness configurations and control isolation due to fragmented layer-wise decision-making. In this paper, we propose a hierarchical orchestration framework with fine-grained coordination for serverless edge clouds, namely Sandwich. To address resource mismatching, we design a fine-grained adaptive readiness configuration mechanism, which extends static cold/warm/hot states into a multi-state framework with dynamically adjustable granularity, enabling precise adaptation to heterogeneous SLOs and workload dynamics. To resolve control isolation, we develop a socialized cross-layer coordination framework that aligns multi-timescale decisions through temporal synchronization, bidirectional decision interaction, and supply-demand coupling across layers. The effectiveness of this hierarchical design is further supported by theoretical approximation bounds and convergence analysis. Experiments on four real-world traces demonstrate significant improvements, achieving up to an 85% increase in cumulative reward, a 14% gain in system performance, and a 7% reduction in startup delay.
Distributed data analytics clusters are widely used to process and store massive datasets. Erasure coding offers a cost-efficient mechanism for achieving data reliability, but incurs substantial bandwidth overhead for failure recovery. Most prior studies on erasure-coded repair focus on minimizing repair bandwidth. However, we observe that erasure-coded repair performance is affected not only by network bandwidth but also by the limited computational capacity under distributed data analytics workloads. We present COAR, a computation-aware erasure-coded repair scheme designed for storage clusters in distributed data analytics environments. COAR exploits the periodic computation patterns of analytics workloads to estimate repair computational capacity via the decoding throughput metric. It strategically assigns and schedules repair tasks by jointly considering computational and network resource states across nodes. Experiments conducted on Alibaba Cloud show that COAR reduces single-chunk, multi-chunk, and full-node repair time by up to $\mathbf{5 9. 8 0 \%, 5 8. 0 7 \%}$, and $\mathbf{5 9. 4 7 \%}$, respectively, compared with existing repair schemes under diverse analytics workloads.
Deep packet inspection (DPI) increasingly struggles to scale with growing network bandwidths and expanding rule sets, with regular expression (regex) matching emerging as a fundamental bottleneck. Conventional regex engines rely on state-based execution, causing processing cost to grow rapidly with rule-set size and limiting scalability even with software or hardware acceleration. In this paper, we present RAGNAROS, an inference-based regex processing system that rethinks regex matching as a pattern recognition problem. RAGNAROS formulates regex matching as a multi-class inference task by modeling each regex as a class in a unified embedding space and performing matching through neural inference rather than per-pattern state traversal. To ensure safe deployment in securitycritical DPI environments, RAGNAROS integrates error-aware confidence analysis with selective fallback to deterministic regex engines. We implement a prototype of RAGNAROS and evaluate it on 10,000 real-world regex patterns from production DPI rule sets. Our results show that RAGNAROS sustains stable 4.6-5 Gbps throughput without match-triggered performance degradation as the rule set scales, while maintaining controlled false positives and effectively near-zero false negatives through selective fallback. These results suggest that RAGNAROS serves as a complementary acceleration layer for existing regex engines, enabling scalable regex processing in modern DPI systems.
