Dual Randomized Smoothing: Beyond Global Noise Variance

Name: Dual Randomized Smoothing: Beyond Global Noise Variance
Uploaded: 2026-09-17T03:59:18.792626+00:00
Duration: 269 s
Description: Researchers from ETH Zürich developed 'Dual Randomized Smoothing: Beyond Global Noise Variance' to address the trade-off between AI accuracy and safety. The method uses a dual-model approach where a variance estimator predicts optimal noise levels for specific inputs, improving certified accuracy on datasets like ImageNet and CIFAR-10.

Researchers from ETH Zürich developed 'Dual Randomized Smoothing: Beyond Global Noise Variance' to address the trade-off between AI accuracy and safety. The method uses a dual-model approach where a variance estimator predicts optimal noise levels for specific inputs, improving certified accuracy on datasets like ImageNet and CIFAR-10.

Open the full episode