Efficient Ensemble Deep Learning for Full Key Recovery in Side-Channel Analysis of Unprotected Ascon Cipher

Abstract Ascon was selected by NIST as the standardized lightweight cryptography algorithm , which makes the leakage behaviour of its implementations a practical concern for Internet-of-Things deployments. Ensemble-based deep learning has recently been shown to break Ascon with very few traces, but reproductions are rarely reported with a statistically careful treatment of traces-to-disclosure (TTD), the number of attack traces after which the correct key byte is ranked first and stays there. TTD directly governs how fast the key can be found: fewer traces mean shorter acquisition time and a faster key search. It is right-censored whenever a key byte is not disclosed within the attack budget. This paper presents an efficient profiled attack with a reproducible methodology on the power traces of an unprotected software Ascon-128 implementation. The attack targets two algebraic outputs of the first-round S-box, y 4 for the key word K 0 and y 2 for K 1 , and uses a per-byte ensemble of five multilayer perceptrons whose log-probabilities are averaged before key scoring. We contribute (i) a complete , byte-wise methodology that recovers all 16 key bytes, (ii) a censoring-aware evaluation using capped medians, a paired Wilcoxon test and a bootstrap confidence interval, and (iii) a contextual comparison with prior Ascon side-channel studies, distinguishing comparable from non-comparable trace counts. The role of AI is decisive for attack effectiveness: instead of assuming a Hamming-weight 1 leakage model and hand-picked points of interest, the neural networks learn the leakage of the S-box output directly from profiling traces, which turns a partly successful classical attack into a complete key recovery. Using 100,000 profiling traces and 30,000 attack traces, both a single MLP and the ensemble place the correct value of every byte at rank 0, whereas non-profiled CPA with 100,000 traces succeeds for only 2 of 16 bytes. Within a 2,000-trace budget the ensemble discloses 13 of 16 bytes versus 10 for a single MLP, and its capped median TTD is 423.5 versus 995.5 traces (Wilcoxon p = 0.020; bootstrap 95% CI of the median difference [10, 802]). We discuss why the K 1 bytes are markedly easier than the K 0 bytes and state the limits of this single-dataset, single-device study.