Continuous spatial keyword query processing has become a critical technique for dynamically tracking user locations and efficiently retrieving relevant objects, with the widespread adoption of mobile devices and location-sensing technologies. To enhance the flexibility of such query processing and reduce operational expenditures, the fog-cloud architecture is increasingly adopted as a mainstream solution. However, its deployment on semi-trusted service infrastructures raises significant privacy concerns. In the paper, we present SKTAQ, a secure continuous spatial keyword query scheme for fog-cloud environments. SKTAQ builds on the KASTree hybrid index, integrating Paillier homomorphic encryption with KTree for secure storage and AssetTree for encrypted query processing. Efficiency is achieved via secure pruning, an optimized range query algorithm, and dynamic region monitoring. Theoretical analysis proves its security and feasibility, while experiments on real and synthetic datasets validate its superior performance.
As the next generation of digital frontier technology, metaverse renders an immersive experience for metaverse users via avatar. As a paradigm of distributed learning, federated learning (FL) facilitates metaverse users' collaborative training of the global model parameter, thereby promoting the metaverse experience. In practical edge-assisted deployments, metaverse intelligence must be delivered under connectivity and resource constraints such as intermittent access, heterogeneous devices, and edge–cloud resource limitations. However, the existing FL algorithms fail to reflect the feature of metaverse, such as dynamic data updating. To make the FL framework closer to the characteristics of the metaverse scenario, we propose a novel personalized FL algorithm, and its originalities are delineated by the following aspects: (1) Users are stringently classified and more nuanced treating in the gradient iterative calculation with gradient sparsification; (2) The personalized FL algorithm considers knowledge transfer toward learning from the pertinent metaverse users, in which the KL difference function is used to calculate the difference between model parameters and hence adaptively reinforce the collaboration among metaverse users with small data heterogeneity; (3) The update strategy of local private data and public data is designed so that the latest updated data is involved in the next episode; (4) Hosting model parameter is defined to knowledge transfer with predicted knowledge coefficient matrix, and it also used to host avatar when the metaverse user is offline. The simulation results demonstrate that the proposed FL algorithm outperforms the state-of-the-art FL algorithms on both the MNIST, FMNIST and CIFAR-10 datasets. We have opened sourced our code in https://github.com/zhulinlijs/Fed learning.
In the context of cloud computing, workflow scheduling refers to the optimized allocation of tasks to virtual machines subject to flow constraints. The optimization varies in terms of the considered performance metrics and is known to be NP-complete. This paper considers the popular variant of optimization where the desired objective is to minimize the cost of executing the tasks subject to a latency constraint. Such a problem is faced in many mission-critical applications, such as disaster early warning systems and autonomous transportation systems, where timeliness is an intrinsic requirement and constitutes a hard deadline. Given the complexity of the problem, heuristic algorithms are pursued. The paper first maps the workflow to a partially ordered set and then uses such a structure to analyze the feasibility of finding a solution. We then utilize such analysis to categorize the design of two sample workflow schedule heuristics and study the performance of multiple popular scheduling algorithms. The contribution of the paper is invaluable for the theoretical assessment of the achievable performance and for guiding algorithm selection in practice.
Multi-access Edge Computing (MEC) has emerged as a promising paradigm that extends cloud computing capabilities to the network edge. It supports real-time and deadline-sensitive applications running on Mobile Devices (MDs) by deploying computation, storage, and communication resources in close proximity to MDs, thereby improving service reliability and responsiveness in heterogeneous MEC systems. Nevertheless, real-time task offloading in heterogeneous MEC systems remains highly challenging due to dynamic wireless conditions, heterogeneous MD capabilities, resource contention among MDs, and stringent deadline constraints under limited MEC resources. As a result, optimal task offloading becomes a crucial yet nontrivial problem in MEC systems. Existing Multi-Agent Deep Reinforcement Learning (MADRL)-based offloading methods typically follow the Centralized Training and Decentralized Execution (CTDE) paradigm, where MDs update their policies based on explicit parameter sharing and aggregated local observations, often leading to suboptimal offloading decisions and excessive computational overhead. To address these limitations, we propose a novel MADRL-based offloading approach, termed TMAT-O, which integrates Centralized Training and Centralized Execution (CTCE) with Task-level Multi-Agent Transformer (TMAT). Technically, the real-time offloading problem is formulated as a Multi-agent Markov Decision Process (MMDP) and solved using the proposed TMAT. By leveraging self-attention and autoregressive sequence modeling, TMAT learns task-level dependencies globally within centralized sequential task decision-making, thereby deriving more effective offloading policies. Extensive simulation results demonstrate that TMAT-O consistently outperforms representative baseline approaches, achieving a minimum reduction of 2.2% in task completion time while consistently lowering deadline-miss rates across diverse MEC configurations.
The autoregressive nature of Large Language Models (LLMs) inference introduces inherent workload uncertainty, leading to severe head-of-line blocking and resource underutilization, particularly in heterogeneous computing networks. Existing LLM inference scheduling paradigms typically rely on large-scale neural predictors, incurring additional overhead and making it difficult to balance prediction cost against task scheduling efficiency. To address these challenges, this paper proposes a novel semantics-aware scheduling framework that predicts LLM inference workloads using LiPER, a lightweight prompt enhanced response length predictor. We design LiPER as a small scale language model by framing response-length estimation as a semantic classification problem. Specifically, LiPER encodes the prompt and directly predicts a discrete length bin with low measured online overhead. Building on bucketized workload priors, we propose LiPER-Sched, a discrete deep reinforcement learning scheduler. By embedding predicted workloads into the system state, LiPER-Sched proactively dispatches tasks across heterogeneous computing nodes to improve load balance and minimize end-to-end latency. Extensive experiments on real world datasets and simulated clusters show that our framework improves over the evaluated rule-based and DRL baselines. LiPER-Sched reduces P99 tail latency by 45% relative to Round-Robin and achieves performance close to an oracle information reference in the evaluated setting.
