MorphStrata Enhances Time-Series Adversarial Robustness with Layer-Specific Perturbations

Time-series forecasting models are often vulnerable to gradient-based adversarial attacks, and existing defense mechanisms frequently involve a trade-off between robustness and computational cost. This challenge is particularly acute in Moving Target Defense (MTD) strategies, where maintaining multiple randomized model instances can substantially increase training overhead. The new approach, MorphStrata, addresses this by proposing a student generation strategy that employs selective, layer-specific stochastic noise injection. This method builds upon the traditional Morphence defense, using a Transformer as the teacher model and perturbing specific architectural blocks to create diverse student models. This structured heterogeneity allows the defense to adapt to varied data distributions and threat models. Evaluations against benchmarks like Jena Climate and Electricity Load Diagrams demonstrate that MorphStrata maintains comparable adversarial RMSE, especially for high-entropy, periodic datasets where it achieves significant improvements over static baselines under various attack types (FGSM, BIM). Crucially, this enhanced robustness comes with less than a 1% increase in training time compared to the Morphence MTD baseline, making it an efficient solution for robust time-series forecasting.