Probabilistic dummy operations inject randomized activity into power traces to blur key-dependent leakage, blunting classical side-channel attacks such as CPA and DPA. We introduce a profiling attack that treats traces as sequences of windows and learns to separate key-dependent computation from dummy activity. A lightweight recurrent sequence classifier is trained on traces from an identical device with dummies disabled, producing a model that scores windows for key-bearing work.
At attack time, the classifier filters dummy-protected traces and the retained windows feed a standard likelihood or correlation-based-key-ranking stage. The key-recovery advantage arises because filtering removes windows with negligible key-dependent leakage, increasing the effective signal-to-noise ratio for classical distinguishers, while the recurrent architecture’s temporal context enables robust detection despite timing jitter and variable dummy density. On a DES implementation with randomized dummy insertion, our method attains rank-0 with substantially fewer traces than CPA, DPA and a tuned CNN, and remains robust under timing jitter (±10 samples), varying dummy rates (p = 0.3–0.7), and low SNR(≤5dB). We report window-level metrics (AUC, ) and key-level success curves (rank vs. traces), with ablations isolating the effects of alignment error and dummy probability. The results demonstrate that probabilistic dummy insertion alone is insufficient against sequence-aware profiling attacks, and that hybrid DL-classical pipelines can outperform both pure classical and pure end-to-end deep learning approaches.