---
title: "Certified Training for Convolutional Perturbations: Formal Safety Guarantees for Vision Models"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.18195v1"
permalink: "https://feed.astralor.com/items/itm_b25335857a25c21136eb5f08"
published: "2026-07-20T17:36:05.000Z"
score: 81
channel: "ai"
tags: ["CIFAR10","Convolutional Perturbations","Provable Robustness","Certified Training","Motion Blur"]
---

# Certified Training for Convolutional Perturbations: Formal Safety Guarantees for Vision Models

## AI commentary

> This paper extends certified training from Lp-norm perturbations to convolutional perturbations, a significant step for robustness research. For teams deploying vision models in the physical world (e.g., autonomous driving, robotics), this work directly tells you how to obtain formal safety guarantees rather than relying on empirical feelings. Though experiments are only on CIFAR10, the method is general and worth following for larger-scale validation like ImageNet.

## At a glance

Vision models are highly susceptible to convolutional perturbations like motion blur from camera shake, risking failures in critical applications. While data augmentation and adversarial training improve empirical robustness, they lack formal guarantees. This paper introduces a novel certified training approach that efficiently encodes convolutional perturbations to train provably robust models. On CIFAR10, it achieves over 80% robust accuracy against moderate motion blur while maintaining comparable standard accuracy, significantly outperforming adversarial training.

## Key points

- Proposes a certified training method for convolutional perturbations (e.g., motion blur), extending provable robustness to such perturbations for the first time.
- Achieves efficient encoding of convolutional perturbations for train-time verifiable bound computation without significant overhead.
- On CIFAR10, achieves over 80% robust accuracy under moderate motion blur while maintaining standard accuracy comparable to vanilla training.
- Outperforms adversarial training significantly in robustness and identifies hidden vulnerabilities not covered by adversarial training.

## Full text

Vision models have been found to be susceptible to perturbations such as motion blur induced at runtime by a shaking camera. This impedes their deployment in critical applications since phenomena such as slightly blurred vision might lead to failures, for example an object detector missing objects. While 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. We introduce a novel Certified Training approach that leverages an efficient encoding of convolutional perturbations to train provably robust models. Our method significantly outperforms Adversarial Training, achieving, for example, over 80% robust accuracy against motion blur of reasonable intensity on CIFAR10 while maintaining comparable standard accuracy.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_b25335857a25c21136eb5f08