MIDSCAN: Investigating the Portability Problem for Cross-Device DL-SCA

Lizzy Grootjen, Zhuoran Liu, Ileana Buhan

Abstract: In deep learning side-channel analysis, a neural network is employed to develop a profile of our target device. Data from a similar dummy device is used to construct the profile. However, when the profiling device differs from the target device, the profile may not be accurate enough for a successful attack. This study examines the effect of manufacturing-induced inter-device discrepancies across 14 identical 32-bit STM32F303 devices. To map the manufacturing discrepancies for these devices, we create a tool called MIDSCAN - Manufacturing-Induced Discrepancies SCAN. Our analysis revealed that the 14 ChipWhisperer 32-bit devices have limited manufacturing discrepancies. Only in a multilayer perceptron setup, the manufacturing discrepancies between the profiling and target devices affected the attack performance. In this case, devices that showed more discrepancies based on correlation and difference than the profiling device need additional traces for a successful attack. Manufacturing discrepancies between profiling and attack devices do not affect attack performance for convolutional neural network architectures. No additional attack traces were needed to perform a successful attack. Our findings indicate that statistical metrics, as implemented by MIDSCAN, can estimate inter-device discrepancies for identical devices. Finally, we found that manufacturing discrepancies are limited for 32-bit STM32F303 ChipWhisperer targets, eliminating the need for additional measures for cross-device deep learning attacks.

External IDs:dblp:conf/lightsec/GrootjenLB25