Investigation of Security Vulnerabilities in NVM-Based Persistent TinyML Hardware

This study investigates vulnerabilities of future generation nonvolatile memory (NVM)-backed persistent TinyML hardware neural networks to side-channel attacks (SCAs) using electromagnetic (EM) analysis methods. We trained three different tinyML models: MobileNet, ResNet, and EfficientNet on three different standard datasets: F-MNIST, CIFAR-10, and MNIST. The trained networks were then mapped on to a custom FPGA-NVM setup for EM-SCA evaluation. We demonstrate that the information about the stored model parameters/weights can be extracted by applying statistical methods on the collected EM emanation data. Further, we demonstrate that the obtained model parametric information can be used for cloning some of the lightweight edge TinyML models with only 0.5%–10% of total training dataset.