往返时间（RTT）：良性信号还是数据中心工作负载的间接窗口？
Round Trip Time: A Benign Signal or an Indirect Window into Datacenter Workloads?
浏览论文内容
中文总结 AI 辅助
该研究针对多租户数据中心的叶脊架构，开发框架利用RTT观测推断共置工作负载，发现间接RTT观测可实现最高97.3%的运行级准确率，仅逻辑隔离无法防止此类信息泄露。
中文摘要 AI 辅助
多租户数据中心网络日益依赖共享的叶脊（leaf-spine）架构，多个租户的流量会经过共同的网络资源。尽管逻辑隔离机制阻止了租户间的直接访问，但共享的拥塞动态仍可能通过可观测的延迟变化，暴露共置工作负载的间接信息。本文中，我们利用沿重叠网络路径收集的RTT观测数据，研究多租户数据中心架构中由共享拥塞行为引发的网络侧信道漏洞。我们开发了一个框架，用于探究工作负载引发的延迟变化是否包含足够可区分的特征，以在实际部署条件下实现工作负载推断。我们的评估表明，间接RTT观测可揭示有意义的工作负载信息，当工作负载引发的拥塞足够可观测时，在跨路径评估下可达到高达97.3%的运行级准确率。这些发现表明，仅靠逻辑网络隔离可能不足以防止现代数据中心基础设施中通过共享拥塞动态造成的信息泄露。
英文摘要
Multi-tenant datacenter networks increasingly rely on shared leaf-spine fabrics, where traffic from multiple tenants traverses common network resources. While logical isolation mechanisms prevent direct access between tenants, shared congestion dynamics may still expose indirect information about co-located workloads through observable latency variations. In this paper, we investigate a network side-channel vulnerability arising from shared congestion behavior in multi-tenant datacenter fabrics using RTT observations collected along overlapping network paths. We develop a framework to explore how workload-induced latency variations contain sufficiently distinguishable signatures to enable workload inference under realistic deployment conditions. Our evaluations show that indirect RTT observations can reveal meaningful workload information, achieving up to 97.3\% run-level accuracy under cross-path evaluation when workload-induced congestion is sufficiently observable. The findings suggest that logical network isolation alone may be insufficient to prevent information leakage through shared congestion dynamics in modern datacenter infrastructures.
