Exhaustive Deep Learning Power Analysis for Secure Block Ciphers and Its Evaluation

<p>In order to realize a sustainable society that integrates cyber and physical space (Society 5.0), it is important to construct secure Cyber-Physical System (CPS). Devices in industrial systems must be energy efficient for environmental protection and low latency for production efficiency. On the other hand, CPS has issues such as huge energy consumption of the entire system and delays caused by communication processing due to frequent network connections of many devices. Lightweight block ciphers are one of the key technologies to solve these problems and improve the confidentiality of highly secret communication data. PRINCE with low latency and Midori128 with low energy operation are both computationally secure. Furthermore, they have been reported to have tamper-resistant circuits with improved security against the threat of power analysis attacks, which use the power consumption of cryptographic operations to guess secret keys. However, deep learning power analysis attacks have been proposed in recent years, focusing on deep learning, which has dramatically improved performance. Therefore, it is very important to evaluate the tamper resistance of PRINCE and Midori128. Against this background, this study examines the threat of deep learning power analysis attack methods against several implementations of PRINCE and Midori128. The proposed method achieves efficient analysis by generating deep learning training data oriented to implementation types and target ciphers. Evaluation experiments using actual equipment showed that the proposed method is able to analyze all partial keys of tamper-resistant circuits, which were difficult to analyze with conventional attack methods.</p>