Guardado en:
Detalles Bibliográficos
Autores principales: Balauca, Stefan, Müller, Mark Niklas, Mao, Yuhao, Baader, Maximilian, Fischer, Marc, Vechev, Martin
Formato: Preprint
Publicado: 2024
Materias:
Acceso en línea:https://arxiv.org/abs/2403.07095
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866912483549118464
author Balauca, Stefan
Müller, Mark Niklas
Mao, Yuhao
Baader, Maximilian
Fischer, Marc
Vechev, Martin
author_facet Balauca, Stefan
Müller, Mark Niklas
Mao, Yuhao
Baader, Maximilian
Fischer, Marc
Vechev, Martin
contents Training neural networks with high certified accuracy against adversarial examples remains an open challenge despite significant efforts. While certification methods can effectively leverage tight convex relaxations for bound computation, in training, these methods, perhaps surprisingly, can perform worse than looser relaxations. Prior work hypothesized that this phenomenon is caused by the discontinuity, non-smoothness, and perturbation sensitivity of the loss surface induced by tighter relaxations. In this work, we theoretically show that applying Gaussian Loss Smoothing (GLS) on the loss surface can alleviate these issues. We confirm this empirically by instantiating GLS with two variants: a zeroth-order optimization algorithm, called PGPE, which allows training with non-differentiable relaxations, and a first-order optimization algorithm, called RGS, which requires gradients of the relaxation but is much more efficient than PGPE. Extensive experiments show that when combined with tight relaxations, these methods surpass state-of-the-art methods when training on the same network architecture for many settings. Our results clearly demonstrate the promise of Gaussian Loss Smoothing for training certifiably robust neural networks and pave a path towards leveraging tighter relaxations for certified training.
format Preprint
id arxiv_https___arxiv_org_abs_2403_07095
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Gaussian Loss Smoothing Enables Certified Training with Tight Convex Relaxations
Balauca, Stefan
Müller, Mark Niklas
Mao, Yuhao
Baader, Maximilian
Fischer, Marc
Vechev, Martin
Machine Learning
Training neural networks with high certified accuracy against adversarial examples remains an open challenge despite significant efforts. While certification methods can effectively leverage tight convex relaxations for bound computation, in training, these methods, perhaps surprisingly, can perform worse than looser relaxations. Prior work hypothesized that this phenomenon is caused by the discontinuity, non-smoothness, and perturbation sensitivity of the loss surface induced by tighter relaxations. In this work, we theoretically show that applying Gaussian Loss Smoothing (GLS) on the loss surface can alleviate these issues. We confirm this empirically by instantiating GLS with two variants: a zeroth-order optimization algorithm, called PGPE, which allows training with non-differentiable relaxations, and a first-order optimization algorithm, called RGS, which requires gradients of the relaxation but is much more efficient than PGPE. Extensive experiments show that when combined with tight relaxations, these methods surpass state-of-the-art methods when training on the same network architecture for many settings. Our results clearly demonstrate the promise of Gaussian Loss Smoothing for training certifiably robust neural networks and pave a path towards leveraging tighter relaxations for certified training.
title Gaussian Loss Smoothing Enables Certified Training with Tight Convex Relaxations
topic Machine Learning
url https://arxiv.org/abs/2403.07095