Sujie Shao

The rapid adoption of Large Language Model (LLM)-based Multi-Agent Systems (MAS) is hindered by severe security risks, where high-connectivity topologies facilitate the explosive propagation of prompt injections and hallucinations. To address this, we propose the Chain-Type Secure Path Reconstruction Method (C-SPRM), a topological defense framework that fundamentally constrains threat lateral movement by transforming arbitrary interaction graphs into linearized logical execution chains based on path decomposition theory. We formulate the topology reconstruction as a Spatio-Temporal Attack Surface (STAS) minimization problem and develop a Graph-to-Sequence Path Decomposition via Reinforcement Learning (G2S-PDR) model to generate optimized execution sequences that balance concurrency risks against latency. Additionally, an adaptive topology virtualization mechanism is introduced to resolve topological bottlenecks in high-degree networks like star or fully connected graphs. Experimental results demonstrate that C-SPRM significantly reduces the effective blast radius of malicious agents and exponentially increases the difficulty of attack propagation, achieving superior system robustness with controllable computational overhead compared to traditional interaction architectures.

Software-defined (SD) honeypots, as dynamic cybersecurity technologies, enhance defense efficiency through flexible resource allocation. However, traditional SD honeypots face latency and jitter issues under network fluctuations, while balancing adjustment costs with defense benefits remains challenging. This paper proposes a DPU-accelerated SD honeypot security service deployment method, leveraging DPU hardware acceleration to optimize network traffic processing and protocol parsing, thereby significantly improving honeypot environment construction efficiency and response real-time performance. For dynamic attack–defense scenarios, we design an adaptive adjustment strategy combining Stackelberg game theory with deep reinforcement learning (AASGRL). By calculating the expected defense benefits and adjustment costs of optimal honeypot deployment strategies, the approach dynamically determines the timing and scope of honeypot adjustments. Simulation experiments demonstrate that the mechanism requires no adjustments in 80% of interaction rounds, while achieving enhanced defense benefits in 20% of rounds with controlled adjustment costs. Compared to traditional methods, the AASGRL mechanism maintains stable defense benefits in long-term interactions, verifying its effectiveness in balancing low costs and high benefits against dynamic attacks. This work provides critical technical support for building adaptive proactive network defense systems.

The integration of AI into power emergency maintenance faces a critical dilemma: centralized training compromises privacy, while standard Federated Learning (FL) struggles with the statistical heterogeneity (Non-IID) of industrial data. Traditional aggregation algorithms (e.g., FedAvg) treat clients solely based on sample size, failing to distinguish between critical fault data and redundant normal operational data. To address this theoretical gap, this paper proposes a Client-Attentive Personalized Federated Learning (PFAA) framework. Unlike conventional approaches, PFAA introduces a semantic-aware attention mechanism driven by “Device Health Fingerprints.” This mechanism dynamically quantifies the contribution of each client not just by data volume, but by the quality and physical relevance of their model updates relative to the global optimization objective. We implement this algorithm within a collaborative cloud-edge-end architecture to enable privacy-preserving, AR-assisted fault diagnosis. Extensive simulations demonstrate that PFAA effectively mitigates model divergence caused by data heterogeneity, achieving superior convergence speed and decision accuracy compared to rule-based and standard FL baselines.

In the field of communication maintenance, Augmented Reality (AR) applications are critical for enhancing operational safety and efficiency. However, deploying the required multimodal models on resource-constrained terminal devices is challenging, as traditional cloud or on-device strategies fail to balance low latency and energy consumption. This paper proposes a Cloud-Edge-End collaborative inference framework tailored to multimodal model deployment. A subgraph partitioning strategy is introduced to systematically decompose complex multimodal models into functionally independent sub-units. Subsequently, a fine-grained performance estimation model is employed to accurately characterize both computation and communication costs across heterogeneous devices. And, a joint optimization problem is formulated to minimize end-to-end inference latency and terminal energy consumption. To solve this problem efficiently, a Hybrid Genetic Algorithm for DNN Partitioning (HGA-DP) evolved over 100 generations is designed, incorporating constraint-aware repair mechanisms and local neighborhood search to navigate the exponential search space of possible deployment combinations. Experimental results on a simulated three-tier collaborative computing platform demonstrate that, compared to traditional full on-device deployment, the proposed method reduces end-to-end inference latency by 70–80% and terminal energy consumption by 81.1%, achieving a 4.86× improvement in overall fitness score. Against the latency-optimized DADS heuristic, HGA-DP achieves 41.3% lower latency while reducing energy by 59.9%. Compared to the All-Cloud strategy, our approach delivers 71.5% latency reduction with only marginal additional terminal energy cost. This framework provides an adaptive and effective solution for real-time multimodal inference in resource-constrained scenarios, laying a foundation for intelligent, resource-aware deployment.

Emerging technologies are revolutionizing power system operation and maintenance. Intelligent state perception is pivotal for stable grid operation, with small object detection technology being vital for identifying minor hazards in power facilities. However, challenges like small object size, low resolution, occlusion, and low confidence arise in small object detection for power operation and maintenance. This paper proposes PyraFAN, a feature fusion method designed for small object detection, and introduces a cloud-edge collaborative inference based smart detection method. This method boosts detection accuracy while ensuring real-time performance. Additionally, a graph-guided distillation method is developed for edge models. By quantifying model performance and task similarity, multi-model collaborative training is realized to improve detection accuracy. Experimental results show that compared with standalone edge models, the proposed method improves detection accuracy by 6.98% and reduces the false negative rate by 19.56%. The PyraFAN module can enhance edge model detection accuracy by approximately 12.2%. Updating edge models via cloud model distillation increases the mAP@0.5 of edge models by 2.7%. Compared to cloud models, the cloud-edge collaboration method reduces average inference latency by 0.8%. This research offers an effective solution for improving the accuracy of deep learning based small object detection in power operation and maintenance within cloud-edge computing environments.

Leveraging Data Processing Units (DPUs) deployed at network interfaces, the DPU-accelerated Intrusion Detection System (IDS) enables microsecond-latency initial traffic inspection through hardware offloading. However, while generating high-throughput alerts, this mechanism amplifies the inherent redundancy and noise issues of traditional IDS systems. This paper proposes an alert correlation method using multi-similarity factor aggregation and a suffix tree model. First, alerts are preprocessed using LFDIA, employing multiple similarity factors and dynamic thresholding to cluster correlated alerts and reduce redundancy. Next, an attack intensity time series is generated and smoothed with a Kalman filter to eliminate noise and reveal attack trends. Finally, the suffix tree models attack activities, capturing key behavioral paths of high-severity alerts and identifying attacker patterns. Experimental evaluations on the CPTC-2017 and CPTC-2018 datasets validate the proposed method’s effectiveness in reducing alert redundancy, extracting critical attack behaviors, and constructing attack activity sequences. The results demonstrate that the method not only significantly reduces the number of alerts but also accurately reveals core attack characteristics, enhancing the effectiveness of network security defense strategies.

In this paper, we propose a power AR emergency communication image recognition method based on end-side multi-source data fusion and edge synergy to solve the problems of real-time, accuracy, and resource constraints in power communication emergency scenarios. Specifically targeting challenges including real-time response requirements under communication-restricted conditions, complex scene variations, and computational resource limitations of wearable devices, the method adopts the cloud–edge–end collaborative architecture: the end-side collects multimodal image data through wearable AR devices and utilizes multi-scale decomposition and reconstruction techniques for preliminary processing and content complementation; the edge-side deploys the EdgeYOLO model, which is based on the Fourier federated learning method to aggregate and optimize the image features in the frequency domain, realizing the dynamic collaboration of multi-devices; and the cloud-side completes the depth of the CloudYOLO model through deep CloudYOLO model completes the deep computation and global optimization to improve the model performance and adaptability. Experimental results on the COCO dataset demonstrate that our method achieves a 15.7% improvement in mAP while maintaining real-time performance, along with a 55.1% reduction in response time.

The development of smart grids makes the demand for time-sensitive service carrying capacity of the power communication network more and more challenging. In recent years, the Power Integrated Communication Networks (PICN), joint multiple wireline or wireless systems, have brought more choices for business access and at the same time brought the complexity of deterministic transmission control. In this paper, we propose a multi-service transmission control method under the PICN. First, the multi-service transmission control architecture and models with time-sensitive constraints are constructed by combining the characteristics of the power business. After that, a routing algorithm based on Bidirectional A∗ (BA) and a flow scheduling algorithm based on Improved Quantum Particle Swarm (IQPS) algorithm are proposed. Simulation results show that the algorithms proposed in this paper can quickly find the optimized flow transmission path and set the Gate Control List (GCL), reducing the flow transmission delay by about 16.5%.