We investigate the Byzantine attack problem within the context of model training in distributed learning systems. While ensuring the convergence of current model training processes, common solvers (e.g. SGD, Adam, RMS...
详细信息
ISBN:
(纸本)9798350366457;9798350366440
We investigate the Byzantine attack problem within the context of model training in distributed learning systems. While ensuring the convergence of current model training processes, common solvers (e.g. SGD, Adam, RMSProp, etc.) can be easily compromised by malicious nodes in these systems. Consequently, the training process may either converge slowly or even *** develop effective secure distributed learning solvers, it is crucial to first examine attack methods to assess the robustness of these solvers. In this work, we contribute to the design of attack strategies by initially highlighting the limitations of finite-norm attacks. We then introduce the seesaw attack, which has been demonstrated to be more effective than the finite-norm attack. Through numerical experiments, we evaluate the efficacy of the seesaw attack across various gradient aggregation rules.
distributed tracing has become a fundamental tool for diagnosing performance issues in the cloud by recording causally ordered, end-to-end workflows of request executions. However, tracing workloads in production can ...
详细信息
ISBN:
(纸本)9798331528690;9798331528706
distributed tracing has become a fundamental tool for diagnosing performance issues in the cloud by recording causally ordered, end-to-end workflows of request executions. However, tracing workloads in production can introduce significant overheads due to the extensive instrumentation needed for identifying performance variations. This paper addresses the trade-off between the cost of tracing and the utility of the "spans" within that trace through Astraea, an online probabilistic distributed tracing system. Astraea is based on our technique that combines online Bayesian learning and multi-armed bandit frameworks. This formulation enables Astraea to effectively steer tracing towards the useful instrumentation needed for accurate performance diagnosis. Astraea localizes performance variations using only 20-35% of available instrumentation, markedly reducing tracing overhead, storage, compute costs, and trace analysis time.
distributed applications running on virtualizationbased systems and cloud computing have become popular solutions, allowing developers to focus on application logic rather than dealing with the complexities of distrib...
详细信息
ISBN:
(纸本)9798350308297
distributed applications running on virtualizationbased systems and cloud computing have become popular solutions, allowing developers to focus on application logic rather than dealing with the complexities of distributedsystems. However, these applications often become increasingly complex, presenting multiple management challenges. To address this issue, software visualization approaches offer valuable solutions by providing real-time insights into resources and their functionalities, offering a comprehensive overview. This study aims to analyze and evaluate existing software visualization tools for distributed applications on the Kubernetes platform. The objective is to comprehensively examine these tools' features, capabilities, and limitations to understand their effectiveness in visualizing complex distributedsystems. Our findings provide valuable insights into the strengths and weaknesses of the available visualization tools, enabling researchers and practitioners to make informed decisions and advancements in software visualization for distributed applications on the Kubernetes platform. Our research identified eight Kubernetes visualization tools, which were examined and compared based on relevant characteristics related to distributed applications and software visualization standards. However, it is worth noting that despite the excellent work done by the community in establishing these first proposals, these tools currently only support, on average, a visualization of 9% of the total resource types available, as mentioned in the official documentation. Therefore, we propose guidelines followed by a synthesized visualization that can guide further research and development in this area. Our study will assist users in selecting the most suitable Kubernetes visualization tool and encourage researchers and the community to explore new approaches in Kubernetes visualization.
Zero Trust security has recently gained attention in enterprise network security. One of its key ideas is making network-level access decisions based on trust scores. However, score-based access control in the enterpr...
详细信息
ISBN:
(纸本)9798350381993;9798350382006
Zero Trust security has recently gained attention in enterprise network security. One of its key ideas is making network-level access decisions based on trust scores. However, score-based access control in the enterprise domain still lacks essential elements in our understanding, and in this paper, we contribute with respect to three crucial aspects. First, we provide a comprehensive list of 29 trust attributes that can be used to calculate a trust score. By introducing a novel mathematical approach, we demonstrate how to quantify these attributes. Second, we describe a dynamic risk-based method to calculate the trust threshold the trust score must meet for permitted access. Third, we introduce a novel trust algorithm based on Subjective Logic that incorporates the first two contributions and offers fine-grained decision possibilities. We discuss how this algorithm shows a higher expressiveness compared to a lightweight additive trust algorithm. Performance-wise, a prototype of the Subjective Logic-based approach showed similar calculation times for making an access decision as the additive approach. In addition, the dynamic threshold calculation showed only 7% increased decision-making times compared to a static threshold.
Surrogate-assisted evolutionary algorithms (SAEAs) have become a popular method to solve data-driven optimization problems (DOPs), which are common in industry. However, with the development of the Internet of Things,...
详细信息
In this paper we propose P4-based Atomic Multicast (P4mCast), a new in-network atomic multicast protocol to support total order guarantees for State Machine Replication (SMR) in cloud-based fault-tolerant and distribu...
详细信息
ISBN:
(纸本)9798350351712;9798350351729
In this paper we propose P4-based Atomic Multicast (P4mCast), a new in-network atomic multicast protocol to support total order guarantees for State Machine Replication (SMR) in cloud-based fault-tolerant and distributed applications. P4mCast builds on in-network computing, applying leader-based consensus to groups of prominent P4 programmable switches in modern data center networks. P4mCast achieves significantly lower latency overhead in the microseconds scale while increasing the throughput one order of magnitude higher compared to state of the art software-based solutions.
Continuous monitoring is a major component of many applications in wireless sensor network (WSN). In these applications, to reduce the communication overhead, some form of data summary or aggregation can be performed....
详细信息
ISBN:
(纸本)9781665495127
Continuous monitoring is a major component of many applications in wireless sensor network (WSN). In these applications, to reduce the communication overhead, some form of data summary or aggregation can be performed. However, performing non-trivial in-network data processing such as finding frequent items, Top-K monitoring, and clustering efficiently are challenging in practice. In this paper, we present Low-Power Distinct Sum (LDS), a distributed in-network data aggregation primitive that performs the sum of unique items in WSN. LDS serves as the underlying primitive that can be used to implement many distributed data processing efficiently. To demonstrate LDS's capabilities, we design and implement a distributed data streaming application with LDS running on Contiki OS. Compared to the baseline algorithm, LDS can reduce the completion time by up to 66%.
作者：
Chen, JunchaoThou, Jian-taoHao, XinyuInner Mongolia Univ
Natl & Local Joint Engn Res Ctr Intelligent Infor Inner Mongolia Engn Lab Cloud Comp & Serv Softwar Inner Mongolia Key Lab Social Comp & Data ProcCo Hohhot Peoples R China
The accuracy of data analysis depends on data quality, and addressing data consistency issues is a key challenge to improve it. Constant Conditional Functional Dependency (CCFD) is an effective approach that ensures d...
详细信息
ISBN:
(纸本)9798350376975;9798350376968
The accuracy of data analysis depends on data quality, and addressing data consistency issues is a key challenge to improve it. Constant Conditional Functional Dependency (CCFD) is an effective approach that ensures data consistency by enforcing bindings of semantically related values, thus providing quality assurance for data analysis and decision-making processes. However, with the growth of data scale, especially the increasing number of data tuples and attributes, existing single-machine CCFD discovery algorithms face issues of low computational efficiency and lengthy computation time. This paper proposes a time-efficient distributed CCFD discovery algorithm (DCCFD). Through the optimization of data preprocessing and index mapping, the data organization structure is enhanced, laying the foundation for the discovery of CCFDs under distributed conditions. The Spark parallel computing framework is used to partition the dataset, which accelerates the parallel loading and processing of data. Additionally, this algorithm ensures accuracy and processing speed when discovering dependencies by efficiently generating frequent itemsets and verifying CCFDs in parallel. Experiments on multiple real datasets show that, especially with the complex Airline dataset, the DCCFD algorithm not only accurately discovers CCFDs, but also reduces the average running time by 75.64% compared with the preCFDMiner algorithm.
The proceedings contain 37 papers. The topics discussed include: few-shot image classification method with label consistent and inconsistent self-supervised learning;a high performance ai-powered cache mechanism for I...
ISBN:
(纸本)9798331506896
The proceedings contain 37 papers. The topics discussed include: few-shot image classification method with label consistent and inconsistent self-supervised learning;a high performance ai-powered cache mechanism for IoT devices;Q-MMT: Qinqiang Opera generation based on multi-track music transformer;research on human combination continuous motion recognition method;multispectral registration algorithm based on fusion of infrared and color images;improved detection of forged and generated facial images based on ResNet-50;intelligent spectrum management for UAV swarm with multiple combat missions: a DQN-based solution;multispectral registration algorithm based on fusion of infrared and color images;and OCDE: adaptive differential evolution algorithm in distributed intelligent systems.
The Internet of Bio-Nano Things (IoBNT) is an innovative field of research located at the intersection of nanotechnology, biotechnology and information and communication technologies. It aims to enable the seamless in...
详细信息
暂无评论