Year: 2025
Type: conference-paper
Abstract: This work introduces a deterministic variant of randomized smoothing for certifying neural network robustness. Instead of relying on multiple stochastic forward passes with Gaussian noise, the proposed method employs symbolic moment propagation to analytically compute the exact mean and covariance of the output distribution in a single pass. Evaluations on MNIST and Fashion MNIST show that the method yields comparable classification accuracy while achieving, on average, 10% larger certified robustness radii. This deterministic framework offers exact, non-probabilistic guarantees that are particularly attractive for safety-critical applications.
Authors: Andrew Mata, Ali ArjomandBigdeli, Stanley Bak
Institution: Stony Brook University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4410140526