Deep Learning-Based Side-Channel Analysis of Lightweight Block Ciphers: A Survey and Gap Analysis Against the AES Baseline
Abstract
Deep learning has advanced profiled side-channel analysis, but for symmetric block ciphers the progress has concentrated almost entirely on the Advanced Encryption Standard (AES). Lightweight block ciphers are now widely deployed in Internet-of-Things and edge devices and have become the subject of formal standardization, but they have received comparatively limited and methodologically fragmented attention. Drawing on a literature search covering 2016 to 2026, this survey places the lightweight setting and the AES baseline side by side along six axes, namely the S-box class space, the target medium, the countermeasure budget, the dataset maturity, the evaluation metrics, and the leakage model. Deep-learning side-channel analysis is then reviewed across the principal lightweight families, followed by a review of the countermeasure landscape under deep-learning attack. The comparison is consolidated into three cross-tabulated matrices spanning per-cipher coverage, countermeasures and their attack outcomes, and dataset usability. The central finding is that no public masked-microcontroller profiling dataset exists for the canonical standalone lightweight block ciphers PRESENT, SIMON, SPECK, Simeck, and GIFT. Among lightweight primitives, public masked-microcontroller profiling datasets exist only for ASCON and the Clyde tweakable block cipher, two designs associated with authenticated encryption. The comparison further shows that no profiled attack has recovered the key of a masked standalone lightweight block cipher. The lightweight studies use the training methodology established on AES only in part. The benchmark gap is the first of three open problems and the precondition for the other two, namely higher-order and masked lightweight attacks and the unattacked surface of code-based masking.