关键词：
Mobile Edge Computing (MEC);Unmanned aerial vehicles (UAVs);Stochastic geometry;Successful uplink communication probability (SUCP)
摘要：
With the increase of computing-intensive and delay-sensitive applications, mobile edge computing (MEC) technology has sprung up. It effectively satisfies the needs of user equipment (UE) for real-time computing resources by placing servers at the edge of the network. However, traditional MEC infrastructures are constrained by their fixed locations and emergency mobility needs. Unmanned aerial vehicles (UAVs) offer an effective solution with low cost, high mobility, and flexible deployment capabilities. In this paper, we propose a region-centric UAV-assisted MEC model, where the whole network space is divided into a set of hexagonal cells of equal area, and all UAVs in the same cell jointly process UE offloading data. Assume all UAVs are assumed to be equipped with independent MEC servers, and both UAVs and terrestrial UE follow independent homogeneous Poisson point process (PPP) distributions. Using the stochastic geometry analysis framework, we derive the successful uplink communication probability (SUCP) of uplink transmission for users. Finally, we compare the simulation results with the theoretical values to verify the model's accuracy and evaluate the influence of key performance parameters on the network performance.
期刊：
Journal of Lightwave Technology,2025年43(18):8525-8537 ISSN：0733-8724
作者机构：
[Qiuyan Yao; Hui Yang] State Key Laboratory of Information Photonics and Optical Communications, Beijing University of Posts and Telecommunications, Beijing, China;[Bowen Bao] School of Computer and Communication Engineering, Changsha University of Science and Technology, Changsha, China
摘要：
Space division multiplexing elastic optical networks (SDM-EONs) is a promising solution to enhance the transmission capacity of optical networks. However, physical layer impairments (PLIs) degrade the service's quality of transmission (QoT). Additionally, service-differentiated spectrum demand leads to spectrum fragmentation, which is further aggravated by strict PLI constraints. Therefore, we propose a routing, core and spectrum allocation algorithm based on dynamic quantitative impairments for low fragmentation (QISEN-LowSF) in SDM-EONs. This algorithm quantitatively analysis multiple impairments and designs restrictive condition to weaken the fragmentation. Specifically, we design a path selection mechanism that takes transmission distance and integration degree of path resources as the cost metrics. Then, we abstract multiple impairment effects in core, and design a core selection mechanism at the cost of impairment estimation and resource integration of each core. When spectrum is allocated to the service, we design a reduced fragmentation spectrum allocation strategy based on the combination of quantified impairment with low-fragmentation. We also optimize the service impairment estimation model in the QoT calculation phase. The results show that QISEN-LowSF performs better than the comparison algorithms under different traffic loads. From the above, the QISEN-LowSF algorithm has superior dynamic flexibility and can reduce blocking probability and the resource fragmentation. Further, it can effectively handle various traffic loads.
摘要：
The rapid development of the Internet has led to the widespread dissemination of manipulated facial images, significantly impacting people's daily lives. With the continuous advancement of Deepfake technology, the generated counterfeit facial images have become increasingly challenging to distinguish. There is an urgent need for a more robust and convincing detection method. Current detection methods mainly operate in the spatial domain and transform the spatial domain into other domains for analysis. With the emergence of transformers, some researchers have also combined traditional convolutional networks with transformers for detection. This paper explores the artifacts left by Deepfakes in various domains and, based on this exploration, proposes a detection method that utilizes the steganalysis rich model to extract high-frequency noise to complement spatial features. We have designed two main modules to fully leverage the interaction between these two aspects based on traditional convolutional neural networks. The first is the multi-scale mixed feature attention module, which introduces artifacts from high-frequency noise into spatial textures, thereby enhancing the model's learning of spatial texture features. The second is the multiscale channel attention module, which reduces the impact of background noise by weighting the features. Our proposed method was experimentally evaluated on mainstream datasets, and a significant amount of experimental results demonstrate the effectiveness of our approach in detecting Deepfake forged faces, outperforming the majority of existing methods.
期刊：
CCF Transactions on High Performance Computing,2025年:1-10 ISSN：2524-4922
通讯作者：
Xiaotian Li
作者机构：
[Xiaoyong Tang; Xiaotian Li; Ronghui Cao] College of Computer and Communication Engineering, Changsha University of Science and Technology, Changsha, China
通讯机构：
[Xiaotian Li] C;College of Computer and Communication Engineering, Changsha University of Science and Technology, Changsha, China
摘要：
Kubernetes is a well-known distributed system for managing containers. It is essential to elect a leader among the replicas to maintain data consistency and coordinate tasks when deploying certain stateful services in a cluster. There are already many leader election algorithms used in distributed systems, but the cost of implementing these algorithms in a Kubernetes cluster is exorbitantly expensive. The existing leader election algorithms in Kubernetes do not take into account the state of the nodes in the election process for distributing the leader, resulting in unbalanced utilization of the cluster and hindering overall cluster performance. This paper proposes an online, resource-aware leader election algorithm to address the aforementioned issues. The algorithm dynamically retrieves the status of cluster nodes to influence the distribution of leaders, ensuring a more balanced allocation of leadership across nodes. This approach helps optimize cluster performance and load balancing. Through experimental comparisons, the algorithm achieves a minimum improvement of 82% in load balancing effectiveness compared to the default and existing improved leader election algorithms, using the coefficient of variation to validate the results.
摘要：
For device-to-device (D2D) communications in the internet of things (IoT), when the direct links between terminals are unavailable owing to obstacles or severe fading, deploying intelligent reflecting surfaces (IRSs) is a promising solution to reconfigure channel environments for enhancing signal coverage and system capacity. In this paper, to improve system spectrum and energy efficiency, a novel full-duplex (FD) D2D communication model with dual IRSs is presented, where two IRSs are deployed closely to two FD transceivers for assisting the exchange of information between them. Given the budget of total transmit power, maximizing the achievable sum-rate of such IRS-assisted FD two-way system is formulated to optimize the precoding at the two transceivers and the phase shifts at the two IRSs. For such a coupled non-convex problem, we decouple it into two subproblems successfully, which can be solved in an alternate manner with low complexity. Simulation results are presented to validate the superior performance of the proposed D2D communication model compared to the existing models and similar optimization schemes.
For device-to-device (D2D) communications in the internet of things (IoT), when the direct links between terminals are unavailable owing to obstacles or severe fading, deploying intelligent reflecting surfaces (IRSs) is a promising solution to reconfigure channel environments for enhancing signal coverage and system capacity. In this paper, to improve system spectrum and energy efficiency, a novel full-duplex (FD) D2D communication model with dual IRSs is presented, where two IRSs are deployed closely to two FD transceivers for assisting the exchange of information between them. Given the budget of total transmit power, maximizing the achievable sum-rate of such IRS-assisted FD two-way system is formulated to optimize the precoding at the two transceivers and the phase shifts at the two IRSs. For such a coupled non-convex problem, we decouple it into two subproblems successfully, which can be solved in an alternate manner with low complexity. Simulation results are presented to validate the superior performance of the proposed D2D communication model compared to the existing models and similar optimization schemes.
关键词：
Servers;Throughput;Scheduling;Admission control;Internet of Things;Scheduling algorithms;Resource management;Optimization;Delays;Time factors;Adaptive time slice;admission control;edge computing;round-robin (RR);task scheduling
摘要：
The rise of multiaccess edge computing (MEC) speeds up mobile user services and resolves service delays caused by long-distance transmission to cloud servers. However, in task-intensive scenarios, edge server processing limitations lead to buffer congestion, increasing latency and reducing Quality of Service (QoS). Furthermore, the challenges of edge server task processing are increased by the varying deadline requirements of different tasks, the time variability of task arrivals, and the real-time fluctuations of the network. In this work, we propose an adaptive slicing-based task admission scheduling strategy (ASTA) to address these issues. ASTA consists of an adaptive time slice adjustment algorithm (ASTA-I) and a task admission scheduling algorithm (ASTA-II). ASTA-I dynamically adjusts time slices based on real-time network conditions and task flow. ASTA-II first adjusts task priorities dynamically by considering factors, such as data volume, deadlines, network conditions, and buffer locations. After that, ASTA-II formulates different scheduling strategies based on changes in task priorities. These strategies are formulated to improve the throughput efficiency of edge servers and enhance the average response speed of tasks. Simulation results show that compared with the existing O2A and OTDS in different scenarios, the proposed ASTA can reduce the average number of waiting requests in the edge server buffer by 19.53%–57.73% and 20.42%–50.26%, and accelerates the average response speed of tasks by about 39.76% and 32.41%.
摘要：
Log analysis, especially log anomaly detection, can help debug systems and analyze root causes to provide reliable services. Deep learning is a promising technology for log anomaly detection. However, deep learning methods need a large amount of training data, which is hard for a newly deployed system to collect sufficient logs. Transfer learning becomes a possible method to solve the problem that can apply the knowledge from a long-term deployed system (source) to a newly deployed system (target). Existing transfer learning methods focus on transferring the knowledge from a source system to a single target system within the same service, in which the source and the target belong to the same service (e.g. operating system, supercomputer, or distributed system). They achieve low performance when applied to multiple target and different services systems because of the obvious differences in log format, syntax, semantics, and component call between different services and the individual training of multiple models for each target system. To tackle the problems, we propose an unsupervised multi-target cross-service log anomaly detection method based on transfer learning and contrastive learning (LogMTC). LogMTC exploits contrastive learning to learn a single model on combined data from the source and multiple target systems, which can fit multiple target systems simultaneously and improve efficiency. LogMTC exploits a hypersphere loss and two contrastive losses to minimize the feature differences crossing different services. Our experiments on two services (supercomputer and distributed system) and three log datasets show that our method is superior to the existing transfer learning methods in the same service, cross-service, and multi-target log anomaly detection. Compared with the best peer accurate transfer learning algorithm LogTAD, LogMTC improves 1.14%-8.28$\%$% F1 score in multi-target transfer and is 1.12-1.22 times faster.
作者：
Abdulmajeed Abdullah Mohammed Mokbel;Fei Yu;Yumba Musoya Gracia;Bohong Tan;Hairong Lin;...
期刊：
复杂系统建模与仿真(英文),2025年5(1):34-45 ISSN：2096-9929
通讯作者：
Yu, F
作者机构：
[Abdulmajeed Abdullah Mohammed Mokbel; Fei Yu; Yumba Musoya Gracia; Bohong Tan] School of Computer and Communication Engineering, Changsha University of Science and Technology, Changsha, China;[Hairong Lin] School of Electronic Information, Central South University, Changsha, China;[Herbert Ho-Ching Iu] School of Electrical, Electronic and Computer Engineering, University of Western Australia, Perth, Australia
通讯机构：
[Yu, F ] C;Changsha Univ Sci & Technol, Sch Comp & Commun Engn, Changsha 410114, Peoples R China.
摘要：
This paper proposes a novel 5D hyperchaotic memristive system based on the Sprott-C system configuration, which greatly improves the complexity of the system to be used for secure communication and signal processing. A critical aspect of this research work is the introduction of a flux-controlled memristor that exhibits chaotic behavior and dynamic responses of the system. To this respect, detailed mathematical modeling and numerical simulations about the stability of the system's equilibria, bifurcations, and hyperchaotic dynamics were conducted and showed a very wide variety of behaviors of great potential in cryptographic applications and secure data transmission. Then, the flexibility and efficiency of the real-time operating environment were demonstrated, and the system was actually implemented on a field-programmable gate array (FPGA) hardware platform. A prototype that confirms the theoretical framework was presented, providing new insights for chaotic systems with practical significance. Finally, we conducted National Institute of Standards and Technology (NIST) testing on the proposed 5D hyperchaotic memristive system, and the results showed that the system has good randomness.
作者机构：
[Zheng Wu; Kehua Guo; Rui Ding] School of Computer Science and Engineering, Central South University, Changsha, 410083, China;[Sheng Ren] School of Computer and Electrical Engineering, Hunan University of Arts and Science, Changde, 41500, China;[Bin Hu] School of Computer and Communication Engineering, Changsha University of Science and Technology, Changsha, 410114, China;[Xiangyuan Zhu] Hunan Provincial Key Laboratory of Network Investigational Technology, Hunan Police Academy, Changsha, 410138, China
通讯机构：
[Kehua Guo] S;School of Computer Science and Engineering, Central South University, Changsha, 410083, China
摘要：
Real-world data are typically long-tailed, causing neural networks to over-fit head classes and underperform on rare tails. We propose Dual-Level Balanced Learning (DBL), an efficient training framework that balances gradients at both the class and instance levels. DBL combines Class-aware Balancing (CB), which corrects class-level imbalance by re-weighting gradients according to prediction bias; Instance-aware Balancing (IB), which alleviates instance-level imbalance by emphasising the learning of hard examples; and a lightweight Cross-Level Collaboration (CC) scheme that harmonises the two losses. By jointly addressing class- and instance-level imbalance, DBL delivers consistent gains across all classes and most individual samples. Extensive experiments on CIFAR10/100-LT, ImageNet-LT, Places-LT, and iNaturalist18 show that DBL sets new state-of-the-art accuracy on all five benchmarks, confirming its robustness to severe long-tailed distributions.
Real-world data are typically long-tailed, causing neural networks to over-fit head classes and underperform on rare tails. We propose Dual-Level Balanced Learning (DBL), an efficient training framework that balances gradients at both the class and instance levels. DBL combines Class-aware Balancing (CB), which corrects class-level imbalance by re-weighting gradients according to prediction bias; Instance-aware Balancing (IB), which alleviates instance-level imbalance by emphasising the learning of hard examples; and a lightweight Cross-Level Collaboration (CC) scheme that harmonises the two losses. By jointly addressing class- and instance-level imbalance, DBL delivers consistent gains across all classes and most individual samples. Extensive experiments on CIFAR10/100-LT, ImageNet-LT, Places-LT, and iNaturalist18 show that DBL sets new state-of-the-art accuracy on all five benchmarks, confirming its robustness to severe long-tailed distributions.
通讯机构：
[Yang, CF ] I;Informat Engn Univ, Key Lab Cyberspace Situat Awareness Henan Prov, Zhengzhou 450001, Peoples R China.
关键词：
Visual place recognition;cross-environment robustness;pre-trained model;feature learning
摘要：
In the Visual Place Recognition (VPR) task, existing research has leveraged large-scale pre-trained models to improve the performance of place recognition. However, when there are significant environmental differences between query images and reference images, a large number of ineffective local features will interfere with the extraction of key landmark features, leading to the retrieval of visually similar but geographically different images. To address this perceptual aliasing problem caused by environmental condition changes, we propose a novel Visual Place Recognition method with Cross-Environment Robust Feature Enhancement (CerfeVPR). This method uses the GAN network to generate similar images of the original images under different environmental conditions, thereby enhancing the learning of robust features of the original images. This enables the global descriptor to effectively ignore appearance changes caused by environmental factors such as seasons and lighting, showing better place recognition accuracy than other methods. Meanwhile, we introduce a large kernel convolution adapter to fine tune the pre-trained model, obtaining a better image feature representation for subsequent robust feature learning. Then, we process the information of different local regions in the general features through a 3-layer pyramid scene parsing network and fuse it with a tag that retains global information to construct a multi-dimensional image feature representation. Based on this, we use the fused features of similar images to drive the robust feature learning of the original images and complete the feature matching between query images and retrieved images. Experiments on multiple commonly used datasets show that our method exhibits excellent performance. On average, CerfeVPR achieves the highest results, with all Recall@N values exceeding 90%. In particular, on the highly challenging Nordland dataset, the R@1 metric is improved by 4.6%, significantly outperforming other methods, which fully verifies the superiority of CerfeVPR in visual place recognition under complex environments.
In the Visual Place Recognition (VPR) task, existing research has leveraged large-scale pre-trained models to improve the performance of place recognition. However, when there are significant environmental differences between query images and reference images, a large number of ineffective local features will interfere with the extraction of key landmark features, leading to the retrieval of visually similar but geographically different images. To address this perceptual aliasing problem caused by environmental condition changes, we propose a novel Visual Place Recognition method with Cross-Environment Robust Feature Enhancement (CerfeVPR). This method uses the GAN network to generate similar images of the original images under different environmental conditions, thereby enhancing the learning of robust features of the original images. This enables the global descriptor to effectively ignore appearance changes caused by environmental factors such as seasons and lighting, showing better place recognition accuracy than other methods. Meanwhile, we introduce a large kernel convolution adapter to fine tune the pre-trained model, obtaining a better image feature representation for subsequent robust feature learning. Then, we process the information of different local regions in the general features through a 3-layer pyramid scene parsing network and fuse it with a tag that retains global information to construct a multi-dimensional image feature representation. Based on this, we use the fused features of similar images to drive the robust feature learning of the original images and complete the feature matching between query images and retrieved images. Experiments on multiple commonly used datasets show that our method exhibits excellent performance. On average, CerfeVPR achieves the highest results, with all Recall@N values exceeding 90%. In particular, on the highly challenging Nordland dataset, the R@1 metric is improved by 4.6%, significantly outperforming other methods, which fully verifies the superiority of CerfeVPR in visual place recognition under complex environments.
作者：
Ming Long;Song Chen;Le-Bing Zhang*;Fei Peng;Dengyong Zhang
期刊：
Multimedia Tools and Applications,2025年84(26):31609-31631 ISSN：1380-7501
通讯作者：
Le-Bing Zhang
作者机构：
[Ming Long; Song Chen; Dengyong Zhang] School of Computer and Communication Engineering, Changsha University of Science and Technology, Changsha, China;[Le-Bing Zhang] School of Computer and Artificial Intelligence, Huaihua University, Huaihua, China;[Fei Peng] Institute of Artificial Intelligence and Blockchain, Guangzhou University, Guangzhou, China
通讯机构：
[Le-Bing Zhang] S;School of Computer and Artificial Intelligence, Huaihua University, Huaihua, China
关键词：
Face De-morphing;Face morphing attack;StyleGAN2;Facial restoration;Feature separation
摘要：
Face morphing attacks pose a significant threat to face recognition systems. In order to solve this problem, several methods for detecting these attacks have been proposed. However, the restoration of the accomplice’s face image remains in its nascent developmental stage. In this paper, we introduce a novel network architecture termed DFS-Net, which leverages double feature spaces (latent code and feature tensor) and a dual-feature separation (DFS) network. DFS-Net is built upon the StyleGAN2 generator. By utilizing the feature vector and feature tensor extracted by the encoder, distinct separation networks are designed. This design facilitates effective feature separation, enabling DFS-Net to separate the identity features of accomplices. The incorporation of the feature tensor enhances the precision in extracting the identity features of the accomplice and thereby elevates the perceptual quality of the restoration image. Experimental results indicate that DFS-Net can effectively restore the accomplice’s face and outperforms previous works in terms of restoration accuracy and image visual quality.
关键词：
Human pose reconstruction;multimodal data fusion;point cloud data processing ultrawideband (UWB) radar;point cloud data processing ultrawideband (UWB) radar;point cloud data processing ultrawideband (UWB) radar
摘要：
This article proposes a multimodal human pose reconstruction method based on 3-D ultrawideband (UWB) radar images and point clouds, aiming to improve the accuracy of human pose estimation through the fusion of radar images and point cloud data. First, a UWB 3-D imaging radar system is designed, which synchronously collects radar echo signals and optical images, constructing a multimodal dataset covering various common actions and different human characteristics. Radar data processing includes azimuth-range 2-D imaging, target locking, local 3-D imaging, discrete sampling, and maximum projection to generate point cloud data and projection images. Optical image processing uses mature methods to reconstruct 3-D poses as pose labels for point clouds and projection images. To achieve multimodal data fusion, the UWB FusionPose network is designed, comprising an image feature extraction network, a point cloud feature extraction network, and a pose reconstruction network. The image feature extraction network is based on the ResNet-18 framework, while the point cloud feature extraction network adopts a pyramid structure. After feature fusion, a multilayer perceptron (MLP) is used to predict human pose information. Additionally, this article explores the impact of fusion parameters on network performance and verifies the effectiveness of the multimodal network through ablation experiments. Experimental results show that this method effectively utilizes radar point cloud data and projection image data to accurately reconstruct the 3-D pose of human targets. This research not only provides a new human pose reconstruction technique but also offers valuable references for the future development of radar imaging technology and multimodal data fusion methods.
摘要：
With the rapid development of Deepfake technology, social security is facing great challenges. Although numerous Deepfake detection algorithms based on traditional CNN frameworks perform well on specific datasets, they still suffer from overfitting due to an over-reliance on localized artifact information. This limitation leads to degraded detection performance across diverse datasets. To address this issue, this study proposes a dual-branch fusion network called LGDF-Net. LGDF-Net uses a dual-branch structure to process the local artifact features and global texture features generated by Deepfake separately, preserving their unique characteristics. Specifically, the local compression branch utilizes a specially designed local compression module (LCM) that allows the network to focus more accurately on key regions of localized artifacts in Deepfake faces. The global expansion branch enhances the analysis of the global facial context through a global expansion module (GEM), which captures image context information and subtle texture features more comprehensively. Additionally, the proposed multi-scale feature extraction module (MSFE) delves into image features at various scales, enriching the extraction of detailed information. Finally, the multi-level feature fusion strategy (MLFF) improves the integration of local and global features through multiple layers, enabling the network to learn the intrinsic connections between these two types of features. A series of experimental validations demonstrate that the proposed scheme outperforms many existing detection networks in terms of accuracy and generalization ability.
摘要：
In the context of transportation cyber-physical systems (T-CPS), backdoor attacks leveraging traffic images have emerged as a significant security threat. As T-CPS increasingly relies on visual information, such as real-time images captured by traffic cameras, for tasks like traffic sign recognition and autonomous driving, the risk of image-based backdoor attacks has grown substantially. Although various detection-based defense techniques have shown some success in identifying backdoored models, they often fail to fully eliminate backdoor effects, leaving residual security risks. To address this challenge, we propose a Frequency-Domain Hybrid Distillation (FDHD) method for backdoor defense, which effectively weakens the association between backdoor triggers and target labels by combining distillation mechanisms in both the frequency and pixel domains. Furthermore, we design a loss function that integrates feature reconstruction with adaptive alignment, enhancing the student network's ability to mimic the teacher network and thereby bolstering the backdoor defense capability. Extensive experiments conducted by FDHD on multiple benchmark datasets against the five latest attacks demonstrate that our proposed defense method effectively reduces backdoor threats while maintaining high accuracy in predicting clean samples. This approach will protect against image-based backdoor attacks in T-CPS and lay the foundation for enhancing future traffic safety.
摘要：
To meet the stringent requirements of industrial applications, modern Ethernet datacenter networks widely deployed with remote direct memory access (RDMA) technology and priority-based flow control (PFC) scheme aim at providing low latency and high throughput transmission performance. However, the existing end-to-end congestion control cannot handle the transient congestion timely due to the round-trip-time (RTT) level control loop, inevitably resulting in PFC triggering. In this article, we propose a Sub-RTT congestion control mechanism called SRCC to alleviate bursty congestion timely. Specifically, SRCC identifies the congested flows accurately, notifies congestion directly from the hotspot to the corresponding source at the sub-RTT control loop and adjusts the sending rate to avoid PFC's head-of-line blocking. Compared to the state-of-the-art end-to-end transmission protocols, the evaluation results show that SRCC effectively reduces the average flow completion time (FCT) by up to 61%, 52%, 40%, and 24% over datacenter quantized congestion notification (DCQCN), Swift, high precision congestion control (HPCC), and photonic congestion notification (PCN), respectively.
通讯机构：
[Feng, CC ] N;Natl Univ Def Technol, Coll Comp Sci, Changsha 410073, Peoples R China.
关键词：
Pricing;Resource management;Servers;Costs;Internet of Things;Optimization;Games;Computational offloading;multiple-access edge computing (MEC);offloading decision;pricing;resources allocation;Stackelberg game
摘要：
Multiaccess edge computing (MEC) is extensively utilized within the Internet of Things (IoT), wherein end-users pay services to meet the latency demands of their respective tasks. The pricing is impacted not solely by the quantity of data offloaded by the user but also associated with the leased computing and communication resources. Nevertheless, prevailing pricing strategies seldom account for the personalized resource requisites during user offloading. In this article, we present an adaptive pricing-oriented approach for concomitant task offloading and resource allocation, considering hybrid resources, comprising two key components. First, we propose a differential pricing framework for communication and computation resources, where the unit price will be influenced by the proportion of resources rented by users. Subsequently, we design a two-stage Stackelberg game model: 1) employing convex optimization theory to mitigate problem intricacies and 2) employing gradient descent to ascertain the potentially optimal price, thus achieving a balance between minimizing user expenses and maximizing server profitability. Simulation outcomes demonstrate that our approach slashes user costs by 23.3% and enhances average server revenue by 65.6% compared to a flat pricing model with a high-user request rate (five user-initiated requests per 100 ms). This maintains server occupancy within 60% to 80%, thereby alleviating user queuing and refining user Quality of Experience (QoE).
摘要：
Packet-level load balancing has shown its massive potential for long in utilizing super high bisection bandwidth of data center network (DCN). This kind of potential, however, has still not been completely transformed into huge performance enhancement of data transmission. The fundamental reason is that packet-level load balancing can fully utilize the parallel paths of underlying physical network, but suffer from the problem of packet disordering transmission, which greatly impairs the flow-level transmission performance of DCN. This paper explores the root cause of performance impairment generated by packet disordering transmission, and proposes $R^{3}$, a solution focusing on “recognizably releasing redundant acknowledgements” as a building block for data center packet-level load balancer. In $R^{3}$'s heart, the source leaf switch perceives the global packet loss information and selectively intercepts the redundant acknowledgement packets, thus avoiding the TCP-driven end-host from experiencing frequent window reductions and unnecessary packet retransmissions. Experimental results of numerous simulation tests and real implementations show that, after integrating $R^{3}$ into the representative data center packet-level load balancing schemes, the transmission performances of both delay-sensitive and throughput-oriented data center flows are significantly improved. Furthermore, $R^{3}$ is merely implemented by switch, leaving the end hosts and the deployed load balancing scheme totally unchanged.
摘要：
In vehicular edge computing (VEC), most tasks require high real-time and energy requirements, but the mobility of vehicles and the difficulty of intelligent computing make it hard to meet these requirements. Due to the fact that most VEC tasks can be decomposed into smaller granularity, based on the dependencies between small subtasks, the repetition of tasks can be reduced, thereby improving task completion rates. In this work, we explore the dependencies of subtasks in different applications and design a two-stage multihop clustering de-duplication offloading (MCDO) mechanism. First, MCDO gives a multihop two-layer clustering (MTLC) algorithm to divide clusters based on similarities between different tasks. Based on this, MCDO further designs a de-duplication logical hierarchical offloading (DLHO). DLHO forms a directed acyclic graph (DAG) of de-duplicated subtasks in each cluster and offloads these subtasks in a logical hierarchical manner. Simulation results show that, compared to existing approaches PC5-GO, FedEdge, and MD-TSDQN, MCDO can achieve a minimum improvement of 15.1% in terms of latency and 20.8% in terms of energy consumption.
摘要：
Large-scale neural networks-based federated learning (FL) has gained public recognition for its effective capabilities in distributed training. Nonetheless, the open system architecture inherent to federated learning systems raises concerns regarding their vulnerability to potential attacks. Poisoning attacks turn into a major menace to federated learning on account of their concealed property and potent destructive force. By altering the local model during routine machine learning training, attackers can easily contaminate the global model. Traditional detection and aggregation solutions mitigate certain threats, but they are still insufficient to completely eliminate the influence generated by attackers. Therefore, federated unlearning that can remove unreliable models while maintaining the accuracy of the global model has become a solution. Unfortunately some existing federated unlearning approaches are rather difficult to be applied in large neural network models because of their high computational expenses. Hence, we propose SlideFU, an efficient anti-poisoning attack federated unlearning framework. The primary concept of SlideFU is to employ sliding window to construct the training process, where all operations are confined within the window. We design a malicious detection scheme based on principal component analysis (PCA), which calculates the trust factors between compressed models in a low-cost way to eliminate unreliable models. After confirming that the global model is under attack, the system activates the federated unlearning process, calibrates the gradients based on the updated direction of the calibration gradients. Experiments on two public datasets demonstrate that our scheme can recover a robust model with extremely high efficiency.
Large-scale neural networks-based federated learning (FL) has gained public recognition for its effective capabilities in distributed training. Nonetheless, the open system architecture inherent to federated learning systems raises concerns regarding their vulnerability to potential attacks. Poisoning attacks turn into a major menace to federated learning on account of their concealed property and potent destructive force. By altering the local model during routine machine learning training, attackers can easily contaminate the global model. Traditional detection and aggregation solutions mitigate certain threats, but they are still insufficient to completely eliminate the influence generated by attackers. Therefore, federated unlearning that can remove unreliable models while maintaining the accuracy of the global model has become a solution. Unfortunately some existing federated unlearning approaches are rather difficult to be applied in large neural network models because of their high computational expenses. Hence, we propose SlideFU, an efficient anti-poisoning attack federated unlearning framework. The primary concept of SlideFU is to employ sliding window to construct the training process, where all operations are confined within the window. We design a malicious detection scheme based on principal component analysis (PCA), which calculates the trust factors between compressed models in a low-cost way to eliminate unreliable models. After confirming that the global model is under attack, the system activates the federated unlearning process, calibrates the gradients based on the updated direction of the calibration gradients. Experiments on two public datasets demonstrate that our scheme can recover a robust model with extremely high efficiency.
摘要：
Despite the evident advantages of variants of UNet in medical image segmentation, these methods still exhibit limitations in the extraction of foreground, background, and boundary features. Based on feature guidance, we propose a new network (FG-UNet). Specifically, adjacent high-level and low-level features are used to gradually guide the network to perceive lesion features. To accommodate lesion features of different scales, the multi-order gated aggregation (MGA) block is designed based on multi-order feature interactions. Furthermore, a novel feature-guided context-aware (FGCA) block is devised to enhance the capability of FG-UNet to segment lesions by fusing boundary-enhancing features, object-enhancing features, and uncertain areas. Eventually, a bi-dimensional interaction attention (BIA) block is designed to enable the network to highlight crucial features effectively. To appraise the effectiveness of FG-UNet, experiments were conducted on Kvasir-seg, ISIC2018, and COVID-19 datasets. The experimental results illustrate that FG-UNet achieves a DSC score of 92.70% on the Kvasir-seg dataset, which is 1.15% higher than that of the latest SCUNet++, 4.70% higher than that of ACC-UNet, and 5.17% higher than that of UNet.
