Short video streaming has become a dominant content format on mobile platforms. Due to frequent user swipes and time-varying network conditions, video chunks may arrive late, leading to playback stalls or unnecessary bandwidth waste. Although multi-path transmission has been explored to accelerate video delivery, existing approaches mainly focus on transport-layer resource scheduling and optimize path utilization in isolation, without coordinating with application-layer video chunk preloading decisions, which limits their effectiveness. In this paper, we propose JUMP, a client-side multi-path adaptive transmission framework for short video streaming that jointly optimizes video chunk preloading and multi-path scheduling. We formulate the video chunk transmission process as a sequential decision-making problem, which is effectively addressed using an Asynchronous Advantage Actor–Critic (A3C) algorithm to learn adaptive joint policies. Experimental results show that, compared with the baseline schemes, JUMP reduces playback stalling by 7.49%–37.58% in the low-bandwidth scenario, improves average bandwidth utilization by 8.33%–39.68% across three bandwidth settings, and achieves Quality of Experience (QoE) gains of 11.72%–36.16%.
The deployment of vision-language models (VLMs) for on-orbit remote sensing is traditionally hindered by the significant computational and energy demands of the models and the restricted hardware capabilities of individual low Earth orbit (LEO) satellites. This paper proposes OrbitServe, a novel energy-aware distributed on-orbit inference framework designed to orchestrate VLM workloads across a dense satellite-cluster formation. By leveraging a tensor-parallel approach tailored for simultaneous lightwave information and power transfer (SLIPT) enabled optical inter-satellite links (OISLs), Goodput@SLO is maximized by (i) offline profiling, and (ii) an epoch-based planner that jointly selects TP degree and effective concurrency under energy constraints. Trace-driven evaluation with an interactive remote-sensing VLM workload shows that OrbitServe significantly improves sustainable serving capacity across diverse orbital energy regimes, and that SLIPT enlarges the feasible operating region only when communication occupancy is non-negligible.
Intent-based networking enables automated network configuration and management through high-level natural language instructions. The core challenge lies in accurately parsing the diverse user requirements into executable formal specifications. Existing approaches either rely on formal languages for efficiency or employ Large Language Models (LLMs) for direct configuration generation, yet both suffer from limited usability and insufficient parsing accuracy. To address this, we propose a DeBERTa-based joint model that explicitly captures the directional and path semantics of key trigger words via a trigger-aware module and introduces a gated interaction mechanism to inject configuration-specification-level semantics into the entity recognition branch, thereby enhancing discriminative capability across semantic roles. We evaluated our model on a self-constructed dataset against various natural language understanding (NLU) baselines and LLMs. Experimental results show that our approach improves the entity recognition F1 score and overall sentence accuracy by approximately 3 percentage points over the best NLU baseline, while achieving significantly lower entity recognition error rates than both traditional NLU models and LLMs.
As ransomware techniques evolve, their threats have extended to critical infrastructures such as the increasingly digitalized smart grid. The smart grid’s highly interconnected communication and control architecture increases system complexity and broadens the attack surface, making it a prominent target for ransomware. Characterized by stealth, rapid propagation, and destructive capability, ransomware poses severe challenges to traditional defenses that struggle to respond promptly and effectively. To address these limitations, this paper proposes an early detection and active defense approach for ransomware in smart grids. The method deploys customized deceptive resources based on behavioral patterns at different attack stages to induce early exposure of abnormal behaviors for rapid identification. A multi-node collaborative defense mechanism allows independent nodes to securely share threat intelligence, enabling coordinated adjustments of defense strategies. Additionally, a permission-driven access control mechanism is incorporated to address the potential bypass issues associated with relying solely on deception-based techniques. The proposed method provides efficient and generalized protection across the ransomware lifecycle, enhancing the overall resilience of smart grid systems.
Speech Analytics Systems are increasingly deployed in customer service and financial applications, where audio is collected on edge devices for centralized recognition and auditing. This workflow introduces substantial privacy risks because speech conveys both sensitive content and biometric identifiers. However, existing protections are often difficult to deploy because they require developers to integrate heterogeneous Digital Signal Processing and deep learning components, manually tune parameters, and continuously balance evolving privacy policies against downstream utility. We introduce PrivSpeech, an adaptive framework that reduces this burden through string-based privacy queries. PrivSpeech compiles high-level objectives into executable pipelines by automatically selecting and composing operators for content masking and identity protection. To balance privacy and intelligibility, we propose Adaptive Semantic-Preserving Voice Anonymization (ASP-VA), which optimizes transformation parameters to minimize identity leakage while preserving downstream utility. Furthermore, PrivSpeech improves efficiency through shared-prefix execution, which merges overlapping computations across concurrent queries. Extensive evaluations demonstrate that PrivSpeech effectively secures sensitive information while maintaining high utility for speech analytics tasks compared to existing baselines.
