Certified Training for Convolutional Perturbations

Forum topic · 小凯 · 2026-07-22

Summary

This paper by Benedikt Brückner and Alessio Lomuscio introduces a certified training approach for neural networks against convolutional perturbations such as motion blur caused by camera shake. While data augmentation and adversarial training can improve empirical robustness of vision models, they offer no formal safety guarantees, making hidden vulnerabilities hard to identify and mitigate. The proposed method leverages an efficient encoding of convolutional perturbations to train provably robust models. Experiments show the approach significantly outperforms adversarial training: on CIFAR10 it achieves over 80% robust accuracy against reasonable-strength motion blur while maintaining comparable standard accuracy. The work targets safety-critical computer vision deployments where perturbations like slight blurring can cause failures, e.g., object detectors missing objects. Paper: arXiv 2607.18195 (cs.CV, cs.LG).

Paper Overview

Field: Computer Vision Authors: Benedikt Brückner, Alessio Lomuscio Published: 2026-07-20 arXiv: 2607.18195 Categories: cs.CV, cs.LG

Abstract

Vision models have been found vulnerable to perturbations such as motion blur caused by camera shake. This hinders their deployment in critical applications, since phenomena like slight visual blurring can lead to failures — for example, an object detector missing objects. Although methods such as data augmentation or adversarial training can improve empirical robustness, they lack formal safety guarantees, making it difficult to identify and mitigate hidden vulnerabilities.

The authors introduce a novel certified training approach that leverages efficient encoding of convolutional perturbations to train provably robust models. The method significantly outperforms adversarial training: for example, it achieves over 80% robust accuracy against reasonable-strength motion blur on CIFAR10 while maintaining comparable standard accuracy.

Key Contributions

Formalizes robustness against convolutional perturbations (e.g., motion blur) rather than only norm-bounded pixel perturbations
Provides an efficient encoding of convolutional perturbations enabling tractable certified training
Demonstrates strong certified performance: >80% robust accuracy on CIFAR10 motion blur with comparable standard accuracy