Profiled Power Analysis Attacks Using Convolutional Neural Networks with Domain Knowledge

Abstract

Evaluation of cryptographic implementations against profiled side-channel attacks plays a fundamental role in security testing nowadays. Recently, deep neural networks and especially Convolutional Neural Networks have been introduced as a new tool for that purpose. Although having several practical advantages over common Gaussian templates such as intrinsic feature extraction, the deep-learning-based profiling techniques proposed in literature still require a suitable leakage model for the implementation under test. Since this is a crucial task, we are introducing domain knowledge to exploit the full power of approximating very complex functions with neural networks. By doing so, we are able to attack the secret key directly without any assumption about the leakage behavior. Our experiments confirmed that our method is much more efficient than state-of-the-art profiling approaches when targeting an unprotected hardware and a protected software implementation of the AES.