Zhongyi Zhai

Large Language Models (LLMs), with their powerful capabilities, are fundamentally transforming society. Cloud-based LLM deployment has drawbacks, including latency, lack of offline functionality, and high long-term costs. Edge computing-based LLM solutions address these issues by offloading inference tasks to edge servers. This paper formulates and optimizes an LLM inference framework that jointly accounts for inference time and predictive quality. Specifically, to address the natural tendency of selfish users for personal utility maximization, we develop a Game-theoretic Offloading Algorithm (GOALIT) to optimize LLM inference offloading. The approach enables distributed users to iteratively adjust their strategies and converge to a Nash equilibrium. Compared to the optimal tree-based search (OT-GAH), the proposed approach yields a 27% increase in token throughput and shortens inference latency by 20%, while maintaining lower perplexity under dynamic system loads. These findings confirm the effectiveness of our approach in resource-constrained edge environments.

Fairness observed for graph neural networks (GNNs) at a conventional decision threshold need not remain stable under stricter high-threshold decision rules, a deployment issue that we term fixed-threshold fairness fragility. To address this problem, we propose HT-FairGNN, a model-agnostic in-processing framework based on high-threshold soft survival-function alignment (HTSFA), which aligns group-wise soft survival functions over an application-relevant high-threshold region using differentiable soft-threshold estimators; we further establish soft-to-hard bounds showing that the HTSFA objective controls empirical hard-threshold Statistical Parity (SP) and Equal Opportunity (EO) disparities up to terms determined by near-threshold score mass and sigmoid approximation error. Across five benchmark datasets, HT-FairGNN reduces average high-threshold SP and EO disparities by roughly two-thirds relative to FairGNN, while predictive-utility outcomes vary across datasets and evaluation metrics. Overall, the results support the effectiveness of HT-FairGNN in improving high-threshold fairness under the evaluated settings, subject to dataset- and metric-dependent utility trade-offs.

Remote sensing image change captioning (RSICC) fuses computer vision and natural language processing to translate visual differences between bi-temporal remote sensing images into interpretable text, with applications in environmental monitoring, urban planning, and disaster assessment. Multimodal Large Language Models (MLLMs) boost RSICC performance but suffer from inefficient inference due to massive parameters, whereas lightweight models enable fast inference yet lack generalization across diverse scenes, which creates a critical timeliness-generalization trade-off. To address this, we propose the Dual-Constrained Transformer (DCT), an end-to-end lightweight RSICC model with three core modules and a decoder. Full-Level Feature Distillation (FLFD) transfers hierarchical knowledge from a pre-trained Dinov3 teacher to a Generalizable Lightweight Visual Encoder (GLVE), enhancing generalization while retaining compactness. Key Change Region Adaptive Weighting (KCR-AW) generates Region Difference Weights (RDW) to emphasize critical changes and suppress backgrounds. Hierarchical encoding and Difference weight Constrained Attention (HDC-Attention) refine multi-scale features via hierarchical encoding and RDW-guided noise suppression; these features are fused by multi-head self-attention and fed into a Transformer decoder for accurate descriptions. The DCT resolves three core issues: lightweight encoder generalization, key change recognition, and multi-scale feature-text association noise, achieving a dynamic balance between inference efficiency and description quality. Experiments on the public LEVIR-CC dataset show our method attains SOTA among lightweight approaches and matches advanced MLLM-based methods with only 0.98% of their parameters.

In mixed traffic scenarios involving both motorized and non-motorized participants, accurately predicting future trajectories of surrounding vehicles remains a major challenge for autonomous driving. Predicting the motion of powered two-wheelers (PTWs) is particularly difficult due to their abrupt behavioral changes and stochastic interaction patterns. To address this issue, this paper proposes an enhanced Social-GAT model with a multi-module architecture for PTW trajectory prediction. The model consists of a dual-channel LSTM encoder that separately processes position and motion features; a temporal attention mechanism to weight key historical states; and a residual-connected two-layer GAT structure to model social relationships within the interaction range, capturing interactive features between PTWs and surrounding vehicles through dynamic adjacency matrices. Finally, an LSTM decoder integrates spatiotemporal features and outputs the predicted trajectory. Experimental results on the rounD dataset demonstrate that our model achieves an outstanding ADE of 0.28, surpassing Trajectron++ by 9.68% and Social-GAN by 69.2%. It also attains the lowest RMSE values across 0.4–2.0s prediction horizons, confirming its superior accuracy and stability for PTW trajectory prediction in mixed traffic environments.

The secret key is stored in an ideal tamper-proof device so that a vehicle can implement a secure authentication with the road-side units (RSUs) and other drivers. However, some adversaries can capture the secret key by physical attacks. To resist physical attacks, we propose a physical-preserving authentication based on a physical unclonable function for vehicular ad hoc networks. In the proposed scheme, a physical unclonable function is deployed on the vehicle and the RSU to provide a challenge–response mechanism. A secret key is only generated by the challenge–response mechanism when it is needed, which eliminates the need to store a long-term secret key. As a result, this prevents secret keys from being captured by adversaries, improving system security. In addition, route planning is introduced into the proposed scheme so that a vehicle can obtain the authentication key of RSUs on its route before vehicle-to-infrastructure authentication, which greatly speeds up the authentication when the vehicle enters the RSUs’ coverage. Furthermore, a detailed analysis demonstrates that the proposed scheme achieves security objectives in vehicular ad hoc networks. Ultimately, when contrasted with similar schemes, the performance assessment demonstrates that our proposed scheme surpasses others in terms of computational overhead, communication overhead and packet loss rate.

In target-oriented multi-agent tasks, agents collaboratively achieve goals defined by specific objects, or targets, in their environment. The key to success is the effective coordination between agents and these targets, especially in dynamic environments where targets may shift. Agents must adeptly adjust to these changes and re-evaluate their target interactions. Inefficient coordination can lead to resource waste, extended task times, and lower overall performance. Addressing this challenge, we introduce the regulatory hierarchical multi-agent coordination (RHMC), a hierarchical reinforcement learning approach. RHMC divides the coordination task into two levels: a high-level policy, assigning targets based on environmental state, and a low-level policy, executing basic actions guided by individual target assignments and observations. Stabilizing RHMC’s high-level policy is crucial for effective learning. This stability is achieved by reward regularization, reducing reliance on the dynamic low-level policy. Such regularization ensures the high-level policy remains focused on broad coordination, not overly dependent on specific agent actions. By minimizing low-level policy dependence, RHMC adapts more seamlessly to environmental changes, boosting learning efficiency. Testing demonstrates RHMC’s superiority over existing methods in global reward and learning efficiency, highlighting its effectiveness in multi-agent coordination.

This paper designs a smart grid energy-saving control system for real-time video streaming, named Eco-VisionGrid. According to the real-time video stream collected by the camera, through the location analysis model and intelligent lighting control model deployed on the edge device, it analyzes the location information of the person and forms the best lighting decision. Preliminary experimental results show that Eco-VisionGrid can effectively implement power regulation and management.