GPUs have been increasingly adopted as hardware accelerators for neural networks (NN). Prior studies show that the unique execution model and many DC-DC units on GPUs make it uneasy to reverse engineer detailed information of encryption algorithms through power side-channel attacks. However, we observe that a few core parameters of NN can be acquired through power profiling when victim and attacker share a GPU. We discuss potential solutions for the power side-channel attacks on GPUs.
展开▼
