Dissecting the Origins of the Power Side-Channel in Deep Neural Networks

In this study, we investigate the vulnerability of Deep Neural Networks (DNNs) to architectural inference via software-based power consumption analysis on GPUs. This work introduces the use of Nvidia-smi, an important tool pre-installed with Nvidia drivers, enabling us to bypass additional software requirements on the target computer. We delve into the distinct power footprints of specific neural networks, look into the units of neural networks, and introduce mitigation techniques. With a widened scope encompassing six different GPU platforms, ranging from physical to the latest high-end cloud-based units, our experiments reveal nuanced insights. This research underscores the effectiveness of power draw patterns in exposing DNN architectures and the critical importance of adopting countermeasures to safeguard intellectual property in the age of advanced machine learning technologies.