Deep Learning Side-Channel Collision Attack

With the breakthrough of Deep Neural Networks, many fields benefited from its enormously increasing performance. Although there is an trend to utilize Learning (DL) for Side-Channel Analysis (SCA) attacks, previous works made specific assumptions attack work. Especially concept template attacks widely adapted while not much attention was paid other strategies. In this work, we present a new methodology, that able exploit side-channel collisions in black-box setting. particular, our performed non-profiled setting and requires neither hypothetical power model (or let’s say many-to-one function) nor details about underlying implementation. While existing DL training metrics distinguish correct key, more efficient by can be applied recover multiple key portions, e.g., bytes. order perform on raw traces instead pre-selected samples, further introduce DL-based technique localize input-dependent leakages masked implementations, associated one byte cipher state case AES. We validated approach targeting several publicly available consumption datasets measured implementations protected different masking schemes. As concrete example, demonstrate how successfully bytes ASCAD dataset with only single trained