FISHI: Fault Injection Detection in Secure Heterogeneous Integration via Power Noise Variation

As Moore's law comes to a crawl, heterogeneous integration-based system-in-package emerges as a promising direction to maintain the speedy rate of performance density improvement of modern integrated circuits by integrating fabricated silicon dies into a unified package. However, hardware security threats such as fault injection attacks present formidable challenges to the protection of on-chip assets. Even worse than a conventional monolithic device, system-in-package (SiP) might introduce malicious chiplets to allow for internal and remote power fault injection attacks due to the obscurity of the semiconductor supply chain. In order to thwart fault injection attacks, we present FISHI which aims to include a root-of-trust chiplet in the SiP to enable run-time system-level power noise variation monitoring capabilities and near-sensor machine learning inference for attack-induced anomaly detection. Specifically, we design a time-to-digital converter to collect power profiles of targeted applications as a reference and create a hardware ML engine accordingly to measure the deviations between the run-time power fluctuations and the golden ones. We prototype our FISHI solution on one of the chiplets in a Xilinx 2.5D FPGA SiP and demonstrate its effectiveness by detecting power fault injection attempts on an AES implementation on the other chiplet.