Revealing Secret Key from Low Success Rate Deep Learning-Based Side Channel Attacks

Non-profiled deep learning-based side channel attacks utilize deep neural networks to extract highly accurate sensitive information. These attacks pose a significant threat to the security of cryptographic devices. Unlike profiled attacks, non-profiled attacks do not require prior knowledge of the target device, making them more versatile. Deep learning algorithms enable attackers to learn complex relationships between side channel signals and secret information, enabling the recovery of cryptographic keys, even the common SCA countermeasure deployed. However, non-profiled DLSCA can not reveal the secret key if the correct key's metric is not clearly distinguished from the incorrect candidates. This paper discusses the mentioned issue of non-profiled DLSCA. Then, a new metric based on the inversion of exponential rank (IER) is proposed to enhance the performance of these attacks. The experimental results show that the proposed technique could reveal the secret subkey even if the partial success rate percentage is only 10% in the ASCAD dataset. Furthermore, when utilizing minimally tuned models and IER metric to execute attacks on the CHES-CTF 2018 data, there is a substantial increase in the percentage of correctly revealed bytes, rising from 62.5% to 93.75%.