DP-SoftShape: Adaptive Differential Privacy via Attention-Guided Sparsification for Time-Series Classification

Local Differential Privacy (LDP) is the de facto protocol for releasing time-series data under formal privacy guarantees. Existing LDP schemes typically add Laplace noise of equal scale at every time step, which tends to drown out the short discriminative patches a classifier depends on. DP-SoftShape spends ϵ unevenly across patches. An attention head scores each shape patch by its contribution to the class label and the Laplace scale at that patch is set in inverse proportion to the score, so noisier perturbation falls on flatter regions while the patches that actually carry the class signal stay close to the original embedding. A uniform-budget version of the same architecture loses most of this gain (Section 4). A Mixture-of-Experts refinement layer downstream recovers features that the noise still disturbs. On 20 UCR datasets at four privacy budgets, DP-SoftShape attains a mean accuracy of 0.722, against 0.565 for ROCKET and 0.553 for Arsenal under the same input-LDP setting.