Currently, manufacturers ensure the security of high-performance embedded devices by employing cryptographic algorithms whose theoretical security has been verified both domestically and internationally. However, even such cryptographic algorithms can...
Currently, manufacturers ensure the security of high-performance embedded devices by employing cryptographic algorithms whose theoretical security has been verified both domestically and internationally. However, even such cryptographic algorithms can suffer from side-channel vulnerabilities due to side-channel information leaked during device operation. Existing profiled side-channel analysis studies have mainly focused on low-performance devices based on single-core microcontroller units(MCUs). In contrast, high-performance devices have additional noise factors such as operating system scheduling, multi-core architectures, and dynamic frequency scaling making the direct application of conventional techniques unsuitable. Moreover, most prior studies, from an academic perspective and for experimental convenience, do not distinguish between the profiling device and the target device. In practical cross-device environments, however, where the profiling device and the target device are different physical devices of the same model, domain discrepancies caused by inter-device variations can significantly degrade analysis performance.
To address these challenges, this study considers a cross-device environment in which device-level discrepancies exist among high-performance devices. We perform deep learning-base side-channel analysis(DLSCA) using electromagnetic(EM) traces acquired from Raspberry Pi 2B platforms equipped with an ARM Cortex-A7-based multi-core processor. To reflect realistic attack scenarios, the default device environment including the operating system, multi-core architecture, and dynamic frequency settings was left unchanged without artificial manipulation.
When the profiling device and the target device differ, domain discrepancies arise due to variations in side-channel acquisition conditions and subtle hardware differences between devices. These discrepancies are a major cause of performance degradation in DLSCA. Device Discrepancy Correction(DDC)[1] mitigates inter-device domain differences by aligning attack data at the input signal level using meand and variance statistics. However, input-level normalization alone is insufficient to fully correct mismatches in feature representations formed within deep learning models. To overcome this limitation, we propose an unsupervised domain adaptation method based on Adaptive Batch Normalization(AdaBN), which realigns batch normalization statistics using attach data, thereby directly matching layer-wise feature distributions to the attack domain.
Experimental results demonstrate that AdaBN enables the guessing entropy to stably converge to zero for all key bytes. In terms of the average normalized total guessing entropy(NTGE), the proposed method achieves approximately a 10.43% performance improvement over the optimal DDC scheme, confirming its effectiveness for cross-device profiled side-channel analysis(CD-DLSCA).