A New Foe in GPUs: Power Side-Channel Attacks on Neural Network

Hyeran Jeon1, Nima Karimian2, Tamara Lehman3
1University of California Merced, 2San Jose State University, 3University of Colorado Boulder


GPUs have become increasingly adopted in the modern system as hardware accelerators in neural networks (NN), however, it introduces new security and privacy challenges. In this work, we propose a GPU hardware scheduling approach to enforce consistent power behavior throughout the NN execution that effectively incapacitate power side-channel attack from reverse engineering hyper-parameters of NN architecture.