The *shuffle model* is a powerful tool to amplify the privacy guarantees of the *local model* of differential privacy. In contrast to the fully decentralized manner of guaranteeing privacy in the local model, the shuffle model requires a central, trusted shuffler. To avoid this central shuffler, recent work of Liew et al. (2022) proposes shuffling locally randomized data in a decentralized manner, via random walks on the communication network constituted by the clients. The privacy amplification bound it thus provides depends on the topology of the underlying communication network, even for infinitely long random walks. It does not match the state-of-the-art privacy amplification bound for the shuffle model (Feldman et al., 2021). In this work, we prove that the output of~$n$ clients' data, each perturbed by an ${\epsilon}_0$-local randomizer, and shuffled by random walks with a logarithmic number of steps, is $( {O} ( (1 - e^{-\epsilon_0} ) \sqrt{ ( e^{\epsilon_0} / n ) \ln (1 / \delta ) } ), O(\delta) )$-differentially private. Importantly, this bound is independent of the topology of the communication network, and asymptotically closes the gap between the privacy amplification bounds for the network shuffle model (Liew et al., 2022) and the shuffle model (Feldman et al., 2021). Our proof is based on a reduction to the shuffle model, and an analysis of the distribution of random walks of finite length. Building on this, we further show that if each client is sampled independently with probability~$p$, the privacy guarantee of the network shuffle model can be further improved to $( {O} ( (1 - e^{-\epsilon_0} ) \sqrt{p ( e^{\epsilon_0} / n ) \ln (1 / \delta ) } ) , O(\delta) )$. Importantly, the subsampling is also performed in a fully decentralized manner that does not require a trusted central entity; compared with related bounds in prior work, our bound is stronger.
翻译:* shuffle 模型 * 是一个强大的工具, 用来扩大 lidireal 模型的隐私保障。 与本地模型中完全分散的保障隐私的方式相反, 洗发模式需要一个中心、 信任的洗发器。 为了避免中央洗发器, 利乌等人( 2022) 最近的工作 提议以分散的方式, 通过客户组成的通信网络随机散行, 以本地随机方式洗发数据 。 因此, 隐私增色约束 取决于 基本通信网络的表层, 甚至无限长的随机行走 。 它与 shulfle 模型( Feldman 和 al. littal ) 的状态更新不匹配 。 lifle 网络( elderman et al. 2021) 运行到 Olittle 模式( litte. slittle mod) 的输出由 $@ little modeal= 20, little man 和 we little view (l_ el_ el_ e) a.