New Framework for Certified Robust Regression Touts Tighter Certificates via Gradients

Randomized smoothing has been a go-to technique for certifying the adversarial robustness of classifiers, but applying it to regression has been an under-explored challenge. Existing regression certificates rely on probabilistic acceptance regions that fail to exploit the local geometry of the function, leading to loose guarantees. Now, a paper accepted at ICML 2026 introduces a higher-order framework that addresses these limitations head-on. The researchers derive a prediction-centered certificate that ensures stability of the smoothed model’s prediction while remaining practically computable at test time. By explicitly incorporating means, variances, and crucially, gradients, the framework produces much tighter robustness certificates than the state-of-the-art alpha-smoothing baseline.

To demonstrate its effectiveness, the team evaluated the method on an MNIST rotation regression task. The results show that utilizing gradient information yields significantly tighter certificates, meaning the model's predictions can be provably robust to larger input perturbations. This breakthrough is particularly important for regression tasks in safety-critical domains like autonomous driving (e.g., steering angle prediction) or medical imaging (e.g., continuous health metrics), where even small adversarial changes could have severe consequences. The work opens the door to more reliable and verifiably robust regression models, pushing the frontier of certified AI beyond classification.