Real-time data sharing at the IoV edge faces high mobility, bursty delays, and constrained RSUs. We present DBG–LB, an edge-deployable loop that couples Dynamic Behavior Graphs (DBG), a time-weighted trust model (TWDT-Trust), a trust-weighted hybrid BFT (TWH-BFT), and budget-bounded incentives. DBG encodes spatio-temporal interaction quality and suppresses correlated evidence via discounting and adaptive decay. TWDT-Trust then fuses direct observations with credibility-controlled recommendations and penalties. TWH-BFT uses a per-round trust snapshot to drive committee sampling and weighted voting, and it supports asynchronous finalize/rollback under heavy-tailed delays. Finally, budget-bounded incentives settle rewards and penalties on chain and feed back to trust and eligibility. In trace-driven emulations with urban/highway mobility, heavy-tailed delays, and collusion/Sybil/replay attacks, DBG–LB sustains sub-second end-to-end latency (median about 0.6 s) and about 200 transactions/s with low malicious inclusion. Compared with PBFT and a PoW-like proxy, it reduces both latency and energy per transaction while maintaining similar or higher finalization. Ablations confirm the necessity of correlation discounts and adaptive decay. Practically, these properties support city-scale V2X deployment by improving road safety and operational efficiency within realistic edge budgets and enabling auditable, policy-compliant collaboration between fleets and authorities.

Visual grounding relies on reasoning between visual and language modalities. Existing multimodal interaction methods struggle to handle complex cross-modal relationships and perform poorly in dynamic scenes. Most paradigms are constrained by single spatial-domain attention, making it difficult to capture global context, long-range dependencies, and balance local and global features. To address these challenges, we propose the Harmonized Spectrum-Gaussian Adaptive Attention Mechanism (HSGAM), a novel mechanism that combines frequency-domain and Gaussian adaptive modulation. HSGAM transforms visual and language features into the frequency domain, overcoming the limitations of spatial-domain self-attention in capturing long-distance dependencies. It also introduces Gaussian adaptive modulation to dynamically adjust feature interactions based on the characteristics of the visual and language modalities. Additionally, we propose the Refinative Discriminative Frequency Network, a feedforward network incorporating enhancement-mitigation and gating mechanisms. Extensive experiments on five benchmark visual grounding tasks illustrate the superiority of our network.

In recent years, learned image compression has witnessed significant advancements. However, many existing learned image compression models predominantly rely on Convolutional Neural Networks (CNNs) and predicting the distribution of latent features to use ever more expensive entropy models, which cannot fully capture detail redundancy and lead to prohibitively slow decoding. To address this challenge, this paper introduces novel image compression models, MGIC and RMGIC, which consist of the Residual Neighborhood-based Attention Module (RNAM), Multi-Scale Guide Entropy Model (MGEM) and Refined Multi-Scale Guide Entropy Model (RMGEM). Specifically, Neighborhood Attention (NA) is combined with global feature learning to construct the RNAM, which can accurately capture the correlation of spatial neighborhoods and is more suitable for image compression. The proposed RNAM is flexible and could work as a plug-and-play component to enhance CNN-based models. Additionally, MGEM employs two mask convolutions with two kernel sizes to leverage multi-scale information for context modeling and eliminating the effect of parallel decoding on model performance. Building upon MGEM, RMGEM incorporates a refined two-stage spatial context modeling technique to enhance the utilization of spatial neighborhoods and achieve more precise entropy prediction. Furthermore, MGIC and RMGIC are derived by integrating RNAM, MGEM, and RMGEM into the variational autoencoder framework. The proposed methods can save more bandwidth and improve the quality of information storage. Extensive experimental evaluations have shown that the proposed method is effective and achieves state-of-the-art performance.

The rising crowdsourced live video streaming (CLVS) poses great challenges to Internet transport scalability, where a broadcaster’s live video is expected to reach thousands and even millions of viewers in real time. To accommodate such huge concurrent video traffic, the de-facto solution is to employ content delivery network (CDN), which distributes the traffic spatially relative to end viewers, using geographically distributed servers. However, our measurement study over a top operational CLVS platform reveals that CDN is not scalable enough, i.e., it loses efficacy, particularly duringbusy timeand leads to tremendous QoE degradation, e.g., 33.3% video bitrate reduction, in comparison to networkidle time. In this work, we propose Spider, which revives the peer-to-peer (P2P) networking principle to extend the scalability of CLVS system. Beyond traditional P2P for elastic data transmission, Spider retrofits P2P to meet the stringent low-latency requirements of CLVS: proposing a “pair-push” streaming mode to tame the excessive signaling latency; designing a QoE-driven peer pairing algorithm to tackle the Internet path variation and CLVS viewer dynamics. We implement, deploy and evaluate Spider in real-world over 20.9 thousand video sessions. Compared to the de-facto CDN solution, Spider achieves remarkable gains, e.g., video stall rate reductions of 52.57%, video quality gains of 8.22%, and even 66% CDN bandwidth saving. The results validate the feasibility and practicability of embracing P2P for low-latency live video communication for the first time.

Maximizing the quality of experience (QoE) for live video streaming is a long-standing challenge. Traditional video transport protocols, represented by a few deterministic rules, can hardly adapt to the highly heterogeneous and dynamic modern Internet environments. Emerging learning-based algorithms have demonstrated the potential to meet above challenge. However, our measurement study reveals an alarming long-tail performance issue: these learning-based algorithms tend to be bottlenecked by occasional catastrophic events due to the built-in exploration mechanisms. In this work, we propose Loki-plus, which improves the robustness of learning-based model by coherently integrating it with a rule-based algorithm. To enable integration at deep feature level, we first reverse-engineer the rule-based algorithm into an equivalent “black-box” neural network, and then devise a transformer-based continual learning model with effective historical feature reservation. Then, we design a network feature-aware fusion mechanism to fuse the two models in a deep manner. We train Loki-plus in a full-fledged live video system through online learning, and evaluate it over massive video sessions, in comparison to state-of-the-art rule-based and learning-based solutions. The results show that Loki-plus improves not only the average but also the tail performance substantially.

The semantic information obtained from large-scale computation in image compression is not practical. To solve this problem, we propose an Attention Aggregation Mechanism (AAM) for learning-based image compression, which is able to aggregate attention map from multiple scales and facilitate information embedding. Based on AAM and Swin Transformer, we design a Multi-scale Self-Attention Image Compression (MSAIC) approach as shown in Figure 1, which is using Multi-scale Swin Transformer blocks (MSTB) and convolutional layers stacks in the down-sampling encoder and up-sampling decoder to the best embed information. The MSTB is made up of convolution layers and three shifted window multi-head self-attention (SW-MSA) blocks, each of which has a different number of heads. In this way, a single MSTB block can aggregate three scales of attention map from tiny, middle and large patches. It can also function as a plug-and-play component to enhance CNN-based models. Meanwhile, we use the channel-wise autoregressive entropy model for the efficient and accurate entropy probability estimation. Extensive experiments show MSAIC is effective and attain the state-of-the-art (SOTA) compression performance, also outperform the SOTA method at the high bit rate.

Maximizing the quality of experience (QoE) for real-time video is a long-standing challenge. Traditional video transport protocols, represented by a few deterministic rules, can hardly adapt to the heterogeneous and highly dynamic modern Internet. Emerging learning-based algorithms have demonstrated potential to meet the challenge. However, our measurement study reveals an alarming long tail performance issue: these algorithms tend to be bottle-necked by occasional catastrophic events due to the built-in exploration mechanisms. In this work, we propose Loki, which improves the robustness of learning-based model by coherently integrating it with a rule-based algorithm. To enable integration at feature level, we first reverse-engineer the rule-based algorithm into an equivalent "black-box" neural network. Then, we design a dual-attention feature fusion mechanism to fuse it with a reinforcement learning model. We train Loki in a commercial real-time video system through online learning, and evaluate it over 101 million video sessions, in comparison to state-of-the-art rule-based and learning-based solutions. The results show that Loki improves not only the average but also the tail performance substantially (26.30% to 44.24% reduction of stall rate and 1.76% to 2.17% increase in video throughput at 95-percentile).