Hardware attacks on ReRAM-based AI accelerators

Resistive random-access memory (ReRAM) is a promising new architecture for artificial intelligence (AI) applications using in-memory computing (IMC). ReRAM-based IMCs deliver 10x better performance, power for AI algorithms. These IMC architectures will be used for AI applications in the future, including for medical, self-driving cars, defense, etc., where security and safety are critical. However, as we show in this work, ReRAM-based IMC architectures are vulnerable to new attacks that are often different from conventional CMOS-based systems. Here, we present voltage-enabled fault-injection, and an aging attack on ReRAM-based AI accelerators. The paper also discusses possible defenses against these attacks and presents open challenges for future work.