Portability Profiling Attacks Based on ACGAN_SDM

In the study of side‐channel attacks, it is typically assumed that leakage traces measured from different devices are similar, allowing analysis and attacks to be conducted using a single device. However, in real‐world attacks, this assumption does not always hold. Variations in manufacturing processes, environmental factors, and other differences across devices can lead to poor generalization of models trained on one device when applied to another. To address the issue of model portability in deep‐learning‐based side‐channel attacks (DL–SCA), this paper proposes the auxiliary classifier generative adversarial network single device model (ACGAN_SDM). This method leverages an auxiliary classifier generative adversarial network to generate synthetic leakage traces that emulate variations expected across different devices, using only traces from a single profiling device. The classifier is then jointly trained on both the real profiling traces and these synthetic multidevice‐like traces. During the attack phase, the classifier processes target leakage traces to predict the probability distribution of secret intermediate values, enabling key recovery through statistical analysis. Through experiments on deep‐learning‐based side‐channel attacks across both single‐device and cross‐device scenarios, we demonstrate that the proposed approach effectively mitigates severe overfitting issues encountered in real‐world attacks, thereby improving model transferability.