Design of Deep Learning Methodology for AES Algorithm Based on Cross Subkey Side Channel Attacks

Recent Deep-Learning Side-Channel Attacks (DLSCAs) utilize networks of neural that have been skilled over part of tracings that merely includes actions relevant to subkeys that were targeted. These attacks are known as “side-channel attacks.” However, due to the restricted number of training traces that are available for deep learning models, such as in the case of the ASCAD database, there is a risk of overfitting occurring. A technique known as “data augmentation” is an example of a data-level method. This technique makes use of additional traces that have been synthetically altered to act as a regularizer and provide deep learning models with a greater capacity for generalization. Within the scope of this investigation, we provide a cross-subkey training approach that acts as an addition to traces. In order to train deep learning models, we use segments for the other 15 subkeys of AES-128 in addition to a trace segment that includes the SBox operation of the target subkey. We show that developing a network model with a mixture of numerous subkeys is superior to the more conventional method of training a network model with a single subkey by using two well-known datasets.