One challenge of the CHES 2018 side channel contest was to break a masked AES implementation. It was impressively won by Gohr et al. by applying ridge regression to obtain guesses for the hamming weights of the (unmasked) AES key schedule, and then using a SAT solver to brute force search the remaining key space. Template attacks are one of the most common approaches used to assess the leakage of a device in a security evaluation. Hence, this raises the question whether ridge regression is a more suitable choice for security evaluation, especially w.r.t. portability. We investigate the feasibility of template attacks to break the presented AES implementation, analyze the leakage of the device, and based on this mount a template attack on hamming weights of the key expansion. We then use classical key search algorithms to recover the AES key. By analyzing the leakage and applying dimension reduction techniques we are able to compress each trace from 650 000
points to only 30 points that are then used to create the templates. Our experimental results indicate that such classical templates achieve similar results compared to ridge regression, and in several cases even slightly outperforming it. According to the organizers, the CTF was aimed to evaluate the concepts of deep learning and classic profiling. Our final conclusion is that the challenge traces are not optimal to settle the question intended, as the leakage is very strong and local. Therefore it is very suitable to apply classical machine learning techniques such as template attacks or ridge regression, and the difficulty in recovering the key is more linked to the resulting key search problem than to the actual attack.