ID 原文 译文
56218 如何构建可靠高效的数据存储系统,成为大数据时代迫切需要解决的问题. How tobuild a reliable and efficient data storage system has become an urgent problem to be solved in the era of bigdata.
56219 相比传统磁盘,非易失主存具有性能高以及字节寻址等优点,这些独特的优势为高效存储系统的构建提供了新的机遇. Compared with traditional disks, non-volatile main memory has the advantages of high performance andbyte-addressable characteristics, and these unique advantages provide new opportunities for the construction ofefficient storage systems.
56220 然而,传统存储系统的构建方式不适用于非易失主存,无法发挥出非易失主存的性能优势,并且容易造成一致性开销高、空间利用率低、编程安全性低等问题. However, the construction of a traditional storage system is not suitable for non-volatilemain memory. It cannot give play to the performance advantage of non-volatile main memory, and it is easyto result in high consistency overhead, low space utilization ratio, and low programming security.
56221 为此,本文分析了基于非易失主存构建存储系统面临的挑战,在系统软件层次分别综述了空间管理机制、新型编程模型、数据结构、文件系统和分布式存储系统等方面的研究进展,并展望了基于非易失主存构建存储系统的未来研究方向. Therefore,this paper analyzes the challenges for the storage system based on non-volatile main memory, and summarizesthe research development of the spatial management mechanism, new programming model, data structure, filesystems, and distributed storage systems. Finally, this paper lists the potential research direction of the storagesystem based on the non-volatile main memory.
56222 联邦机器学习系统由于能够在多方之间训练联合模型而无需各方共享训练数据,因此在学术界和工业界都获得了越来越多的关注和应用. Federated machine learning systems have gained more and more attention and popularity in bothacademia and industry because they can obtain a shared model among multiple parties without explicitly sharingthe training data.
56223 与传统的机器学习框架相比,这类系统被认为具有保护数据隐私的良好潜力. Such a system is believed to have a good potential of protecting data privacy compared with thetraditional machine learning frameworks.
56224 另一方面,训练阶段攻击是一种通过故意扰动训练数据,从而希望在测试时操纵相应的学习系统预测行为的攻击方法. On the other hand, training time attacks are a procedure of purposefullymodifying training data, hoping to manipulate the behavior of the corresponding trained system during test time.
56225 例如, DeepConfuse是最近的一种高效生成对抗训练数据的方法,展示了传统监督学习范式在此类攻击下的脆弱性. DeepConfuse, for instance, is one recent advance in generating adversarial training data with high efficiency.
56226 在本文中,作者扩展了DeepConfuse方法,将其应用在联邦机器学习框架中. Inthis work, we extend the DeepConfuse framework so that it can be used in federated machine learning.
56227 这是首次针对联邦学习系统的训练阶段攻击. This isthe first training time attack for a federated learning system.