The case for adding privacy-related offloading to smart storage

Abstract

It is important to ensure that personally identifiable information (PII) is protected within large distributed systems and is used only for intended purposes. Achieving this is challenging and several techniques have been proposed for privacy-preserving analytics, but they typically focus on the end hosts only. We argue that future storage solutions should include, in addition to emerging compute offload, also privacy-related operators. Since many privacy operators, such as perturbation and anonymization, take place as the very first step before other computations, query offload to a Smart Storage device might be only feasible in the future if privacy-related operators can also be offloaded. In this work we demonstrate that privacy-preserving operators can be implemented in hardware without reducing read bandwidths. We focus on perturbations and extend an FPGA-based network-attached Smart Storage solution to show that it is possible to provide these operations at 10Gbps line-rate while using only a small amount of additional FPGA real-estate. We also discuss how future faster smart storage nodes should look like in the light of these additional requirements.