Enhancing Cryptographic Security Through a Machine Learning Model for Early Detection of Side-Channel Attacks

Abstract

This paper presents a profiling method of side-channel leakage of AES algorithm desynchronized traces based on the ASCAD dataset using deep learning devices. The model relies on a convolutional neural network (CNN) design to automatically derive useful features of the time-series measurements of power, surmounting the noise and keep-up issues.
As shown by experiment, the proposed model has a high training accuracy of 95.4% and a validation accuracy of 95.1% when the number of epochs reached 60, which means that the model has achieved good generalization and adaptability to irregular and noisy data. The model manages to find pattern relations that are related to the secret key bytes, which shouldn't be underestimated as to the possibility of deep learning to improve side-channel analysis even in very difficult circumstances.
These results imply the relevance of machine learning to cryptographic security and inform about the efficiency of attacks.