On-die Differential Sensing for Monitoring and Analysis of Dynamic Computing Environments

With increasing popularity of cloud FPGAs and future multi-tenant usage, security threats are posed including on-die side channel attacks, fault injection or denial of service caused by power wasting circuits. Monitoring FPGA circuit activity with sensors is a common technique used by attackers and defenders. Combining data from multiple sensors makes it possible to pinpoint power wasting circuits, but with limited precision. In this paper we demonstrate how a network of sensors and differential analysis can together increase robustness of monitoring, making it possible to determine the location of target circuits even if their power consumption is relatively low. The sensors are ring oscillator (RO) or time-to-digital converter (TDC) circuits that collect timeseries information of the local voltage at different FPGA points. The sensor data and knowledge of when target circuits are active, enables the differential analysis technique that boosts the sensitivity beyond that of prior works. We analyze how the performance varies with sensor coverage and circuit parameters and show how a pre-characterization procedure can further improve accuracy over uncalibrated analysis.