Tolerance Evaluation Against Deep Learning Side-Channel Attack on AES in Automotive Microcontroller With Uncertain Leakage Model

Side-channel attacks (SCAs) pose a threat to cryptographic security because they potentially recover cryptographic secret-keys by analyzing electromagnetic noise from cryptographic ICs. Interest in automotive security has been growing in recent years, resulting in a demand for SCA tolerance evaluation of automotive ICs. In the current paper, we implemented AES128 in software on the Renesas RH850 series microcontroller and demonstrated leading SCA methods such as correlation power analysis (CPA) and deep learning SCA (DL-SCA). CPA needs leakage models and is therefore more powerful against unprotected AES in the ATmega series and the STM series. However, our results show that CPA fails to recover all bytes of the cryptographic secret-key even with 800,000 leakage traces, whereas DL-SCA succeeds in recovering all bytes of the key. These results differ from those with ATmega and STM series, indicating that information leakage from the microcontroller is sufficient to recover the keys, but that the conventional leakage model is not valid for CPA.