The prediction of network latency with partial measurements is of importance for ever-increasing personal devices to ensure their Quality of Service (QoS). However, the current matrix-factorization-based efforts, as a promising paradigm, for network latency prediction have failed to intelligently exploit inherent factors hidden in networks to accurately infer the unknown network latency. Furthermore, it is more complicated to execute extensive network measurements on pervasive personal devices due to unstable communication environments. To alleviate these problems, in this paper, a novel accurate network latency prediction (DALP) solution via Deep Reinforcement Learning (DRL) is proposed for personal devices under cost-aware sampling. Specifically, we first alternately implement cost-aware latency measurement based on temporal correlation, and model it as a network latency matrix, in which unmeasured and missing elements need to be inferred. In order to achieve accurate prediction performance, the DRL-based Matrix Factorization with Double Weights (DWMF) is designed to exploit the potential network factors and multiple rules of matrix factorization, which can be alternatively executed, to minimize the prediction errors. Furthermore, an angle-loss-based reward strategy is designed to enhance the quality of model training. Simulation results on real-world datasets illustrate that DALP outperforms the previous approaches with quicker convergence and lower prediction errors.

Recent self-supervised stereo matching methods have made significant progress. They typically rely on the photometric consistency assumption, which presumes corresponding points across views share the same appearance. However, this assumption could be compromised by real-world disturbances, resulting in invalid supervisory signals and a significant accuracy gap compared to supervised methods. To address this issue, we propose SMFormer, a framework integrating more reliable self-supervision guided by the Vision Foundation Model (VFM) and data augmentation. We first incorporate the VFM with the Feature Pyramid Network (FPN), providing a discriminative and robust feature representation against disturbance in various scenarios. We then devise an effective data augmentation mechanism that ensures robustness to various transformations. The data augmentation mechanism explicitly enforces consistency between learned features and those influenced by illumination variations. Additionally, it regularizes the output consistency between disparity predictions of strong augmented samples and those generated from standard samples. Experiments on multiple mainstream benchmarks demonstrate that our SMFormer achieves state-of-the-art (SOTA) performance among self-supervised methods and even competes on par with supervised ones. Remarkably, in the challenging Booster benchmark, SMFormer even outperforms some SOTA supervised methods, such as CFNet.

Diffusion models can effectively generate intricate images by gradually refining noise into detailed visual data. Diffusion inversion maps a real-world image to a feature space to achieve image attribute manipulation. Existing diffusion inversion techniques cannot effectively integrate semantic and visual information, resulting in insufficient control precision. This paper introduces the Cross-Modal Contrastive Inversion (CMCI) method to enhance single-image attribute manipulation for pre-trained diffusion models. CMCI learns cross-modal conditioning embedding through contrastive regularization, combining visual and textual information to guide denoising. Our method helps diffusion models generate images semantically matched with textual descriptions while preserving the integrity of the source image. CMCI can be applied to single-video attribute manipulation tasks by capturing both static features and dynamic temporal changes through our contrastive regularization. Comprehensive experiments are performed, and the results suggest that CMCI outperforms competing inversion methods in terms of both inversion accuracy and manipulation capabilities.

Integrated sensing and communication (ISAC) provides an effective solution for supporting both high-rate data transmission and environmental perception in connected and automated vehicles (CAVs). However, in urban environments, buildings frequently hinder signal propagation between the base station (BS) and vehicles, thereby compromising link reliability and weakening the effectiveness of radar-based sensing. To mitigate these challenges, intelligent reflecting surface (IRS) technology has been proposed to enhance CAV operations by improving the signal propagation environment. This study focuses on the ISAC system implemented in IRS-assisted CAV scenarios. In this system, the IRS facilitates the simultaneous transmission of communication and radar signals. By jointly optimizing the active beamforming matrix at the BS and the passive beamforming matrix at the IRS, the sum rate achievable by vehicular users is maximized, while ensuring compliance with the BS’s power budget, maintaining required radar detection quality, and adhering to the IRS phase configuration limits. In light of these considerations, an approach, namely Adaptive Optimization for Sensing and Forming (AOSF), is proposed, which employs the semidefinite relaxation (SDR) and successive convex approximation (SCA) algorithms to optimize the active beamforming matrix at the BS and employs the fractional programming algorithm to optimize the passive beamforming matrix at the IRS. The method decomposes the non-convex optimization problem into a series of manageable subproblems and employs an iterative solving approach to gradually refine the results and converge toward a satisfactory solution. Simulation results indicate that the proposed approach significantly improves achievable sum rate performance and reveals the trade-offs between communication and sensing.

Federated Learning (FL) enables privacy-preserving collaboration across distributed edge devices but faces severe challenges in cross-device settings due to stringent memory and computation constraints. While quantization offers a natural remedy, existing techniques primarily target inference acceleration and fail to improve training efficiency. Consequently, current quantized FL frameworks only mitigate communication overhead while suffering from memory inefficiency during on-device training. Moreover, they exhibit Heterogeneity-induced optimization instability, as devices with varying bit-widths jointly participate in federated learning, introducing inconsistent quantization noise that hinders convergence and degrades final accuracy. In this work, we propose AquaFed, a memory-efficient Ascending quantized Federated learning framework tailored for heterogeneous edge environments. We first introduce Memory-Efficient Quantization (MEQ), which quantizes both weights and activation caches during training to drastically reduce peak memory usage and memory I/O. To address instability arising from heterogeneous precision, we theoretically establish the Quantization Noise Equivalence (QNE) Hypothesis: Quantization error exhibits a behavior analogous to the stochastic noise induced by a large learning rate in quantized federated learning. Guided by QNE, AquaFed adopts an ascending quantization schedule, allowing low-bit devices to lead early coarse-grained learning, while high-bit devices refine the model in later stages for stable convergence. Extensive experiments on vision and language benchmarks demonstrate that AquaFed consistently delivers superior accuracy, generalization, and memory efficiency compared to state-of-the-art methods. On CIFAR-10 with ResNet-18, AquaFed achieves a 20% accuracy improvement and 39% memory saving over the best existing framework.

Edge computing can potentially improve user Quality of Experience (QoE) in ever-increasing video streaming by caching popular videos. However, most state-of-the-art client-side ABR algorithms, such as MPC and Pensieve, ignore the presence of a cache on the video delivery path, leading to unsatisfying video QoE. Thus, in this paper, we present MPC2, a novel cache-aware ABR algorithm for cache-based adaptive video streaming, which can optimally combine throughput, buffer occupancy, and cache state based on model predictive control (MPC). Facing the problem of stale bandwidth measurement, MPC2 proposes a bandwidth interpolation scheme based on the features of edge-assisted video streaming, introducing no additional overhead. Moreover, MPC2 further proposes Bitrate Combination Search algorithm based tabu search (BCS-tabu) to resolve the practical deployment issues, which greatly reduces the computation delay. To evaluate the performance of MPC2, we compare MPC2 with the related state-of-the-art works using trace-driven simulation under various real-world traces. The results demonstrate that MPC2 outperforms the best state-of-the-art work, a deep reinforcement learning-based cache-aware ABR, achieving superior optimality and stability in cache-based video streaming systems.

This paper proposes TrustEvo, a novel hi erarchical reinforcement learning framework for task offloading in the Social Internet of Things (SIoT), aiming to achieve an effective balance between efficiency and trust in ubiquitous computing environments. SIoT is considered as a key paradigm to enable decentralized and socially aware ubiquitous computing, where devices act as both task consumers and service providers. To address the challenges of dynamic environments, heterogeneous resources, and hybrid decision spaces, TrustEvo formulates task offloading as a hierarchical decision-making problem. Specifically, TrustEvo decomposes the offloading process into 3 coupled levels, jointly optimizing the offloading type, target node selection, and offloading ratio. This design enables efficient exploration of both discrete and continuous decision spaces while capturing the intrinsic structure of offloading decisions. TrustEvo explicitly integrates trust evolution into the decision-making loop. By modeling trust as a dynamic state variable driven by task execution out comes, TrustEvo establishes a closed-loop interaction between system performance and node reliability, adaptively balancing the multi-objective optimization of latency minimization, success rate maximization and trust value maxi mization under uncertain dynamic environments. Extensive experiments based on real-world datasets demonstrate that TrustEvo method consistently outperforms baseline approaches. TrustEvo achieves lower andmorestablelatency, higher average trust levels, and robust task success rates. These results validate its effectiveness as a lightweight, adaptive, and trust-aware solution for next-generation SIoT enabled ubiquitous computing systems.

The exponential proliferation of short video services has imposed severe strain on mobile backhaul links, necessitating intelligent proactive caching at the network edge to guarantee user Quality of Experience (QoE). Although proactive caching offers a promising avenue to alleviate congestion, its effectiveness is often constrained by the challenge of capturing rapid shifts in user interests and the intricate diffusion of content across social networks. To bridge this gap, we propose SC-STRL, a Social-Cognitive Spatiotemporal Reinforcement Learning framework that integrates temporal interest evolution with social structure analysis. Specifically, SC-STRL models user community influence to identify key opinion leaders and utilizes a Spatiotemporal Encoding Architecture involving a Temporal Interest Evolution Network (TIEN) and a Social Structure Embedding Module (SSEM) to capture dynamic demands. Building on these insights, we further devise a specialized Double Deep Q-Network (DDQN) agent augmented with Prioritized Experience Replay (PER), featuring a novel social-interactive reward mechanism explicitly designed to align caching decisions with content propagation dynamics. Extensive experiments on the KuaiRec dataset demonstrate SC-STRL's superiority, achieving performance improvements of 12.02% over the state-of-the-art benchmarks at a cache capacity of 400.

The convergence of 5G and Time-Sensitive Networking (TSN) offers a powerful foundation for enabling ultra-reliable, low-latency, and deterministic communication across a wide spectrum of emerging applications. While prior surveys primarily concentrate on industrial automation, this work presents, to the best of our knowledge, the first comprehensive survey that systematically investigates the potential of integrated 5G-TSN in four critical domains: industrial automation, intelligent transportation systems, smart energy systems, and digital health. We begin by providing a detailed background and comparative analysis of 5G and TSN technologies, highlighting recent advancements and complementary capabilities. The paper then presents a deep dive into an integrated 5G-TSN system architecture, with particular focus on time synchronization, traffic scheduling, QoS mapping, and cross-domain resource coordination. Building on this technical foundation, we introduce a structured, application-specific analysis that maps the communication requirements, challenges, and domain-specific integration strategies to the corresponding enabling 5G-TSN mechanisms. Finally, we synthesize key technical challenges such as interoperability, end-to-end synchronization, heterogeneous QoS alignment, and unified security, and propose targeted research directions to support the practical, scalable deployment of integrated 5G-TSN for emerging applications.

Radio frequency (RF) signals have gained widespread adoption in intelligent perception systems due to their unique advantages, including non-line-of-sight propagation capability, robustness in low-light environments, and inherent privacy preservation. However, their substantial data volumes, generated by the dual-polarization direction characteristic, result in significant challenges to data storage and transmission. To address this, we propose the first end-to-end deep dynamic RF signal compression (DRFC) framework, which primarily focuses on exploiting cross-directional correlation in dynamic RF signals. The proposed framework incorporates four key innovations: (1) a mask-guided RF motion estimation module that leverages Doppler shifts and electromagnetic noise characteristics to identify regions of significant motion using a threshold-based mask, significantly improving motion estimation accuracy; (2) a cross-directional RF motion entropy model that utilizes cross-directional RF motion latent priors to refine the probability distribution for motion entropy coding; (3) a cross-directional RF context mining module that predicts RF contexts from temporal and cross-directional reference signals, adaptively fusing these contexts with confidence maps to maximize complementary information utilization; and (4) a cross-directional RF contextual entropy model that incorporates cross-directional RF contextual latent priors to optimize contextual entropy modeling. Experimental results demonstrate the superiority of our framework over existing codecs. Our DRFC framework achieves significant bitrate savings on benchmark datasets, establishing a strong baseline for future research in this field.

This paper proposes a secure space-air-ground in tegrated network (SAGIN) transmission scheme in which an unmanned aerial vehicle (UAV) equipped with reconfigurable holographic surfaces (RHS) antennas to tackle the challenges of ultra-long distances in ground-space (G2S) links and eaves dropping risks in wireless channels. Leveraging the holographic beamforming principle, the UAV can relay confidential G2S signals while transmitting interference as artificial noise toward eavesdroppers to counter interception. The closed-form expressions of the secrecy transmission rate for SAGIN are first derived. Then, an optimization strategy is designed under onboard energy consumption constraints, which jointly optimizes the beam radiation amplitude, 3D trajectory and power allocation, aiming to maximize the average secrecy rate (ASR) while satisfying the quality of service for legitimate users. A block coordinate descent algorithm is employed to decouple the original non convex problem into four subproblems, and feasible solutions are obtained through successive convex approximation. Simulation results validate the advantages of the proposed scheme, with the ASR significantly outperforming baseline schemes, such as those without holographic beamforming, horizon-only trajectory deployment, and without power control. The variations of the ASR with RHS element size and UAV transmit power are also analyzed. This research provides new theoretical insights and technical solutions for secure communication in SAGIN.

Low-latency data sensing and transmission is critical for many city-level applications like traffic incident management to mitigate congestion and enhance road safety. Vehicular crowdsensing (VCS) emerges as a powerful paradigm to provide real-time traffic sensing services from points-of-interest (PoIs) by leveraging the collaboration of unmanned ground vehicles (UGVs) and unmanned aerial vehicles (UAVs). In this paper, we first introduce two novel metrics: sensing capability-aware age-of-information ($s$AoI) and latency-weighted data collection ratio, to measure the data freshness and amount under the condition of non-uniform status packet size, respectively. We propose an auto-regressive sequential multi-agent deep reinforcement learning framework called “A2G-MADRL”, which consists of an interaction-aware heterogeneous vehicular graph convolution network (HVGCN) for feature extractions, and a dynamically ordered masked policy generator (DOMPG) for coordinating UAVs and UGVs. Extensive experiments on two real-world datasets in KAIST and Roma demonstrate that A2G-MADRL significantly reduces the attained $s$AoI and improves latency-weighted data collection ratio, outperforming seven baselines when varying the number of UAV-UGV pairs, data generation speed in a timeslot, and the number of communication channels.

Traditional unimodal sensing faces limitations in accuracy and capability, and its decoupled implementation with communication systems increases latency in bandwidth-constrained environments. Additionally, single-task-oriented sensing systems fail to address users' diverse demands. To overcome these challenges, we propose a semantic-driven integrated multimodal sensing and communication (SIMAC) framework. This framework leverages a joint source-channel coding architecture to achieve simultaneous sensing, decoding, and transmission of sensing results. Specifically, SIMAC first introduces a multimodal semantic fusion (MSF) network, which employs two extractors to extract semantic information from radar signals and images, respectively. MSF then applies cross-attention mechanisms to fuse these unimodal features and generate multimodal semantic representations. Secondly, we present a large language model (LLM)-based semantic encoder (LSE), where relevant communication parameters and multimodal semantics are mapped into a unified latent space and input to the LLM, enabling channel-adaptive semantic encoding. Thirdly, a task-oriented sensing semantic decoder (SSD) is proposed, in which different decoded heads are designed according to the specific needs of tasks. Simultaneously, a multi-task learning strategy is introduced to train the SIMAC framework, achieving diverse sensing services. Finally, experimental simulations demonstrate that the proposed framework achieves diverse and higher-accuracy sensing services.

We introduce UCS-Bench, a dataset spanning 170+ hours of egocentric visual observations with 7K+ timestamped questions for diagnosing User-centric Continual Spatial intelligence in egocentric video streams. UCS-Bench targets a new problem that emphasizes dynamic spatial reasoning, long-term memory, and their alignment with users' real-time locations. We propose DirectMe, a framework that incrementally constructs and maintains a structured spatial memory from streaming egocentric observations. DirectMe enables robust tracking and recall of object locations, all relative to user's movement over time. By tightly coupling visual perception with memory updates and spatial reasoning, our approach supports long-horizon queries that require recalling interactions, resolving viewpoint-induced ambiguities, and adaptation to dynamic scenes. Our experiments show that DirectMe significantly improves the spatial reasoning of leading multimodal LLMs; it also surpasses many spatial-aware and long streaming video models. We hope our benchmark and solution will advance spatial intelligence research for egocentric AI assistants. Data and code will be released.

This paper is concerned with autonomous aerial vehicle (AAV) video coding and transmission in scenarios such as aerial search and monitoring. Unlike existing methods of modeling AAV video source coding and channel transmission separately, we investigate the joint source-channel optimization (JSCO) issue for video coding and transmission. Particularly, we design eight-dimensional delay-power-rate-distortion models in terms of source coding and channel transmission and characterize the correlation between video coding and transmission, with which a JSCO problem is formulated. Its objective is to minimize end-to-end distortion and AAV power consumption by optimizing fine-grained parameters related to AAV video coding and transmission. This problem is confirmed to be a challenging sequential-decision and non-convex optimization problem. We therefore decompose it into a family of repeated optimization problems by Lyapunov optimization and design an approximate convex optimization scheme with provable performance guarantees to tackle these problems. Based on the theoretical transformation, we propose a Lyapunov repeated iteration (LyaRI) algorithm. Both objective and subjective experiments are conducted to comprehensively evaluate the performance of LyaRI. The results indicate that, compared with its counterparts, LyaRI achieves better video quality and stability performance, with a 47.74% reduction in the variance of the obtained encoding bit.

Being capable of realizing smart and controllable channel environments, the reconfigurable intelligent surface (RIS) shows great potential to improve both the spectral and energy efficiencies. Due to the multiplicative fading effect of reflective RIS links, the capacity gain induced by passive RIS (PRIS) is limited in communication scenarios with blocked or very weak direct links. This work introduces an active RIS (ARIS) to a multiuser multi-input multi-output (MIMO) broadband communication system with orthogonal frequency division multiplexing (OFDM), aiming at enhancing the transmission rate. Specifically, a sum rate optimization problem is formulated based on a practical broadband reflection model of the ARIS. An algorithm for addressing a suboptimal subcarrier allocation, the optimal precoding matrix of the base station (BS), and the suboptimal reflecting coefficients of ARIS is proposed. Specifically, a channel attenuation coefficient-based Hungarian algorithm is used to fastly address a suboptimal subcarrier allocation. The optimal precoding matrix of the BS is presented explicitly, where water-filling form solution is given for two special cases. The weighted minimum mean square error method associated with a penalty method is used to find a suboptimal reflecting coefficients of the ARIS, which well handle the coefficient coupling across subcarriers inherent in the practical broadband RIS. Numerical results validate the convergence and effectiveness of the proposed algorithm. It is shown that by using the proposed algorithm, the sum rate of the system can be enhanced significantly.

Autonomous vehicles must accurately predict the trajectories of surrounding vehicles to ensure safety and efficiency in complex traffic environments. However, existing approaches often neglect the impact of driving style variations, limiting their ability to model diverse driver behaviors. This paper proposes a Spatial-Temporal Transformer considering Driving Style (DS-STT) for accurate and personalized vehicle trajectory prediction. The model constructs a spatial-temporal traffic graph (STTG) from vehicle trajectories and employs a Transformer-based encoder with Spatial Self-Attention (SSA) and Temporal Convolution (TC) to capture inter-vehicle interactions and temporal dependencies. A Style Attention Module (SAM) is further introduced to integrate driving style, motion state, and interaction features, enabling adaptive prediction for different driving behaviors. Experimental results on the HighD and ApolloScape datasets demonstrate that DS-STT consistently outperforms baseline models, achieving up to 81.58% lower Root Mean Squared Error(RMSE) on the HighD dataset and 22.51% lower Final Displacement Error (FDE) on the ApolloScape dataset. These results confirm the superior accuracy, robustness, and long-term prediction capability of DS-STT, highlighting the significance of incorporating driving style in enhancing the safety and reliability of autonomous driving systems.

With the remarkable progress in neural P-frame video coding, neural B-frame coding has recently emerged as a critical research direction. However, most existing neural B-frame codecs directly adopt P-frame coding tools without adequately addressing the unique challenges of B-frame compression, leading to suboptimal performance. To bridge this gap, we propose novel enhancements for motion compression and temporal fusion for neural B-frame coding. First, we design a fine-grained motion compression method. This method incorporates an interactive dual-branch motion auto-encoder with per-branch adaptive quantization steps, which enables fine-grained compression of bi-directional motion vectors while accommodating their asymmetric bitrate allocation and reconstruction quality requirements. Furthermore, this method involves an interactive motion entropy model that exploits correlations between bi-directional motion latent representations by interactively leveraging partitioned latent segments as directional priors. Second, we propose a selective temporal fusion method that predicts bi-directional fusion weights to achieve discriminative utilization of bi-directional multi-scale temporal contexts with varying qualities. Additionally, this method introduces a hyperprior-based implicit alignment mechanism for contextual entropy modeling. By treating the hyperprior as a surrogate for the contextual latent representation, this mechanism implicitly mitigates the misalignment in the fused bi-directional temporal priors. Extensive experiments demonstrate that our proposed codec achieves an average BD-rate reduction of approximately 10

Edge intelligence enhances the computational capabilities of resource-limited devices by offloading inference tasks to edge servers. Traditional methods either execute the entire model on the device, resulting in slow inference, or fully offload it to the server, incurring communication delays and privacy risks due to raw data transmission. Model partitioning addresses these challenges by splitting the model for execution on both the device and edge server, transmitting only intermediate inference results. However, current model partitioning methods lack consideration of device mobility, resulting in reduced inference efficiency and task interruptions. To address these limitations, we introduce MobiSplit, a novel mobility-aware framework that dynamically partitions inference models between resource-constrained devices and edge servers. MobiSplit adapts to real-time device mobility, fluctuating network conditions, and computational constraints to minimize inference latency and energy consumption while ensuring robust task execution. Additionally, we propose a distributed auction-based algorithm that empowers edge devices to autonomously determine optimal partitioning and offloading strategies in a scalable and adaptive manner. Extensive simulations demonstrate that MobiSplit enhances inference efficiency, achieving a 60% latency reduction and a 20% energy consumption decrease compared to the best-performing baseline across diverse edge scenarios.

Recent self-supervised stereo matching methods have made significant progress, but their performance significantly degrades under adverse weather conditions such as night, rain, and fog. We identify two primary weaknesses contributing to this performance degradation. First, adverse weather introduces noise and reduces visibility, making CNN-based feature extractors struggle with degraded regions like reflective and textureless areas. Second, these degraded regions can disrupt accurate pixel correspondences, leading to ineffective supervision based on the photometric consistency assumption. To address these challenges, we propose injecting robust priors derived from the visual foundation model into the CNN-based feature extractor to improve feature representation under adverse weather conditions. We then introduce scene correspondence priors to construct robust supervisory signals rather than relying solely on the photometric consistency assumption. Specifically, we create synthetic stereo datasets with realistic weather degradations. These datasets feature clear and adverse image pairs that maintain the same semantic context and disparity, preserving the scene correspondence property. With this knowledge, we propose a robust self-supervised training paradigm, consisting of two key steps: robust self-supervised scene correspondence learning and adverse weather distillation. Both steps aim to align underlying scene results from clean and adverse image pairs, thus improving model disparity estimation under adverse weather effects. Extensive experiments demonstrate the effectiveness and versatility of our proposed solution, which outperforms existing state-of-the-art self-supervised methods. Codes are available at https://github.com/cocowy1/RoSe-Robust-Self-supervised-Stereo-Matching-under-Adverse-Weather-Conditions.