Machine Learning-Based Side-Channel Analysis on the Advanced Encryption Standard

Hardware security is essential in keeping sensitive information private. Because of this, it’s imperative that we evaluate the ability of cryptosystems to withstand cutting edge attacks. Doing so encourages the development of countermeasures and new methods of data protection as needed. In this thesis, we present our findings of an evaluation of the Advanced Encryption Standard, particularly unmasked and masked AES-128, implemented in software on an STM32F415 microcontroller unit (MCU), against machine learning-based side-channel analysis (MLSCA). 12 machine learning classifiers were used in combination with a side-channel leakage model in the context of four scenarios: profiling one device and key and attacking the same device with the same key, profiling one device and key and attacking a different device with the same key, profiling one device and key and attacking the same device with a different key, and profiling one device and key and attacking a different device with a different key. We found that unmasked AES-128 can be very vulnerable to this form of attack and that masking can be applied as a countermeasure to successfully prevent attacks in 2 out of the 4 tested scenarios. In addition to providing our experimental results on the following pages, we also plan to release a public GitHub repository with all of our collected side-channel data along with sample analysis code shortly after the time of writing this. We hope that doing so will allow for complete reproducibility of our results and encourage future research without the need for purchasing hardware equipment