UC San Diego

Cloud-FPGAs are an attractive option for accelerating consumer computation without the need to purchase a multi-thousand dollar device. The greatest strength of these devices, their reconfigurability, comes at a cost in terms of opening new attack vectors.

We first consider an attractive cloud-FPGA model that has garnered great commercial and academic interest for reducing costs and maximizing utilization: the virtualization of cloud-FPGA resources, called multi-tenancy. However, side-channel leakage poses a major security threat in multi-tenant FPGA environments. A tenant can instantiate a signal timing sensor that measures minute changes in the power distribution network and infer information about co-tenant computation. This work presents the Tunable Dual-Polarity Time-to-Digital Converter (TDC)—a signal timing sensor with three dynamically tunable parameters: the sample duration, clock phase, and frequency.

Returning to the existing cloud-FPGA model, we present, to the best of our knowledge, the first remote measurement of bias temperature instability, a type of transistor degradation, on a commercial cloud-FPGA platform. We repurpose the same on-fabric TDC testing mechanism as before. A study is provided demonstrating this bias effect within the FPGA routing, characterizing its relationship to the number of transistors in the underlying tested element, and exploring its elastic nature, on three different architectures: PYNQ-Z2, ZCU102, AWS F1. We present a novel attack vector that leverages this effect in cloud-FGPAs, where a malicious user can extract secrets from previous user’s computation.