Detection of Crucial Power Side Channel Data Leakage in Neural Networks
Amjed Abbas Ahmed, Mohammad Kamrul Hasan, Nurhizam Safie Mohd Satar, Nazmus Shaker Nafi, Azana Hafizah Mohd Aman, Shayla Islam, Saif Aamer Fadhil
Abstract
Neural network (NN) accelerators are now extensively utilized in a range of applications that need a high degree of security, such as driverless cars, NLP, and image recognition. Due to privacy issues and the high cost, hardware implementations contained within NN Propagators were often not accessible for general populace. Additionally with power and time data, accelerators also disclose critical data by electro-magnetic (EM) sided channels. Within this study, we demonstrate a side-channel information-based attack that can successfully steal models from large-scale NN accelerators deployed on real-world hardware. The use of these accelerators is widespread. The proposed method of attack consists of two distinct phases: 1) Using EM side-channel data to estimate networking's underlying architecture; 2) Using margin-dependent, attackers learning actively in estimating parameters, notably weights. Deducing the underlying network structure from EM sidechannel data. Inferring the underlying network structure from EM sidechannel data. Experimental findings demonstrate that the disclosed attack technique can be used to precisely retrieve the large-scale NN via the use of EM side-channel information leaking. Overall, our attack shows how critical it is to conceal electromagnetic (EM) traces for massive NN accelerators in practical settings.