Imperceptible Misclassification Attack on Deep Learning Accelerator by Glitch Injection

The convergence of edge computing and deep learning empowers endpoint hardwares or edge devices to perform inferences locally with the help of deep neural network (DNN) accelerator. This trend of edge intelligence invites new attack vectors, which are methodologically different from the well-known software oriented deep learning attacks like the input of adversarial examples. Current studies of threats on DNN hardware focus mainly on model parameters interpolation. Such kind of manipulation is not stealthy as it will leave non-erasable traces or create conspicuous output patterns. In this paper, we present and investigate an imperceptible misclassification attack on DNN hardware by introducing infrequent instantaneous glitches into the clock signal. Comparing with falsifying model parameters by permanent faults, corruption of targeted intermediate results of convolution layer(s) by disrupting associated computations intermittently leaves no trace. We demonstrated our attack on nine state-of-the-art ImageNet models running on Xilinx FPGA based deep learning accelerator. With no knowledge about the models, our attack can achieve over 98% misclassification on 8 out of 9 models with only 10% glitches launched into the computation clock cycles. Given the model details and inputs, all the test images applied to ResNet50 can be successfully misclassified with no more than 1.7% glitch injection.