Adaptive Randomized Smoothing with Certified Robustness for Mitigating Tabular Adversarial Attacks

Deep learning models for tabular classification remain susceptible to adversarial perturbations—small, structured changes to input features that can induce incorrect predictions. While various defenses have been proposed, randomized smoothing has emerged as a certifiable and theoretically grounded technique for building robust classifiers. However, standard randomized smoothing can increase false positives and degrade accuracy on clean, non-perturbed inputs, limiting practical utility in real-world applications. To address this limitation, we introduce a novel methodology that integrates entropy-scaled smoothing during training and margin-confidence proxy-based resampling during certification. This enhanced framework produces a certifiably robust classifier that retains resilience to adversarial perturbations while substantially lowering false positive rates. We empirically evaluate our adaptive randomized smoothing method on eight attacks, two classifiers (a deep neural network and TabNet) and across three benchmark datasets. We validate the effectiveness of our approach in both binary and multi-class classification settings, demonstrating its applicability across different task complexities. In addition to standard adversarial settings, we rigorously test our defense against constrained, realistic attack scenarios to ensure practical robustness. Our results demonstrate improved resilience and reduced false positives, while maintaining reasonable operational efficiency with a controllable efficiency-accuracy tradeoff; fewer samples accelerate certification but may slightly reduce accuracy.