Scholarlink

<p>Deep learning, when integrated with a large amount of training data, has the potential to outperform machine learning in terms of high accuracy. Recently, privacy-preserving deep learning has drawn significant attention from the research community. Different privacy notions, in the context of deep learning, include privacy of data provided by data owners and privacy of parameters and/or hyperparameters of the underlying deep neural network. Federated learning is a popular privacy-preserving execution environment where data owners participate in learning the parameters collectively without leaking their respective data to other participants. However, federated learning suffers from certain security/privacy issues: as each participant has access to the architecture of the whole neural network, a malicious participant can leak the architecture to a third party. In this paper, we propose a variant of split learning, that we call <Emphasis FontCategory="SansSerif">Split-n-Chain</Emphasis>, where the layers (and respective parameters) of the neural network are split among several distributed nodes such that the whole network is not available to any single node. <Emphasis FontCategory="SansSerif">Split-n-Chain</Emphasis> achieves several privacy properties: data owners need not share their training data with other nodes, and no individual node has access to the parameters and hyperparameters of the deep neural network (except those of the respective layer it holds). Moreover, <Emphasis FontCategory="SansSerif">Split-n-Chain</Emphasis> exploits blockchain to audit the computation done by different (possibly malicious) nodes. We implement <Emphasis FontCategory="SansSerif">Split-n-Chain</Emphasis> and experiment with the same in two settings: using a single local machine as well as a group of distributed machines connected through a LAN. Our experimental results show: <Emphasis FontCategory="SansSerif">Split-n-Chain</Emphasis> is efficient, in terms of time required to execute different phases, and the training loss trend in <Emphasis FontCategory="SansSerif">Split-n-Chain</Emphasis> is similar to that for the same deep neural network when implemented in a monolithic fashion.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Split-n-Chain: Privacy-preserving multi-node split learning with blockchain-based auditability

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Mukesh Sahani,
  • Binanda Sengupta

摘要

Deep learning, when integrated with a large amount of training data, has the potential to outperform machine learning in terms of high accuracy. Recently, privacy-preserving deep learning has drawn significant attention from the research community. Different privacy notions, in the context of deep learning, include privacy of data provided by data owners and privacy of parameters and/or hyperparameters of the underlying deep neural network. Federated learning is a popular privacy-preserving execution environment where data owners participate in learning the parameters collectively without leaking their respective data to other participants. However, federated learning suffers from certain security/privacy issues: as each participant has access to the architecture of the whole neural network, a malicious participant can leak the architecture to a third party. In this paper, we propose a variant of split learning, that we call Split-n-Chain, where the layers (and respective parameters) of the neural network are split among several distributed nodes such that the whole network is not available to any single node. Split-n-Chain achieves several privacy properties: data owners need not share their training data with other nodes, and no individual node has access to the parameters and hyperparameters of the deep neural network (except those of the respective layer it holds). Moreover, Split-n-Chain exploits blockchain to audit the computation done by different (possibly malicious) nodes. We implement Split-n-Chain and experiment with the same in two settings: using a single local machine as well as a group of distributed machines connected through a LAN. Our experimental results show: Split-n-Chain is efficient, in terms of time required to execute different phases, and the training loss trend in Split-n-Chain is similar to that for the same deep neural network when implemented in a monolithic fashion.