Soft Error Reliability Assessment of TinyML Algorithms on STM32 Microcontroller

Ultra-low-power and low-cost microcontrollers are converging with the recently popular concept of Tiny Machine Learning (TinyML) to run such algorithms on memory and latency-constrained devices. Although they are frequently used in autonomous drives such as unmanned aerial vehicles (drones), their reliability is often overlooked. In this study, we analyze the soft-error reliability of various TinyML algorithms executing on an STM32L4 series microcontroller. We implemented a low-level fault injector that can inject faults with timer interrupts in the form of a random bit flip in registers/memory during inference. A total of 24,794 tests were conducted across Motion Detection, Image Recognition, and Visual Wake Word as TinyML algorithms and a matrix multiplication application. Our experimental results revealed that soft errors can substantially affect such algorithms, with vulnerabilities differing across registers and memory with different crash and silent data corruption (SDC) rates. These findings u nderscore t he importance o f t ailoredfault tolerance strategies for enhancing the reliability of TinyML models.