International Association
for Cryptologic Research

Profiling side-channel analysis is an essential technique to assess the security of protected cryptographic implementations by subjecting them to the worst-case security analysis. This approach assumes the presence of a highly capable adversary with knowledge of countermeasures and randomness employed by the target device. However, black-box profiling attacks are commonly employed when aiming to emulate real-world scenarios. These attacks leverage deep learning as a prominent alternative since deep neural networks can automatically select points of interest, eliminating the need for secret mask knowledge. Nevertheless, black-box profiling attacks often result in non-worst-case security evaluations, leading to suboptimal profiling models.

In this study, we propose modifying the conventional black-box threat model by incorporating a new assumption: the adversary possesses a similar implementation that can be used as a white-box reference design. We create an adversarial dataset by extracting features or points of interest from this reference design. These features are then utilized for training a novel conditional generative adversarial network (CGAN) framework, enabling a generative model to extract features from high-order leakages in protected implementation without any assumptions about the masking scheme or secret masks. Our framework empowers attackers to perform efficient black-box profiling attack that achieves (and even surpasses) the performance of the worst-case security assessments.