Music to My Ears: Turning GPU Sounds into Intellectual Property Gold

In this paper, we introduce an acoustic side-channel attack that extracts crucial information from Deep Neural Networks (DNNs) operating on GPUs. Utilizing a Micro-Electro-Mechanical Systems (MEMS) microphone with an extensive frequency range, we demonstrate that the distinct sounds produced during DNN operations can inadvertently reveal significant details about the network's architecture. Through extensive experimentation with a variety of neural networks from the ImageNet competition, we validate the efficacy of this novel attack vector. Our contributions include a detailed methodological framework for capturing and analyzing acoustic data, empirical validation using prominent ImageNet models, and a comprehensive sensitivity analysis to assess the attack's robustness under varying conditions. This research not only uncovers a previously unexplored vulnerability in neural network security but also provides a foundation for developing more robust defense mechanisms against such innovative side-channel attacks.