We propose the first certified defense tailored for time-series data under Dynamic Time Warping (DTW) distance.

We propose MIBP-Cert, a certified training method that uses mixed-integer bilinear programming to compute tight robustness guarantees under complex training-time perturbations. Our approach stabilizes training and improves certified accuracy.

We propose AdaptDel, a randomized smoothing technique with adaptable deletion rates, achieving higher certified accuracy for sequence classification under edit distance perturbations, outperforming previous methods on diverse NLP tasks.