Entity · paper

Certified Training for Convolutional Perturbations

paperactiveprovisionalcertified-training-for-convolutional-perturbations-6c874233·1 events·first seen Jul 21, 2026

Aliases: Certified Training for Convolutional Perturbations

Co-occurring entities

More like this (12)

quantization-aware training PC Layer: Polynomial Weight Preconditioning for Improving LLM Pre-Training Dynamic Short Convolutions Improve Transformers Training-Free Looped Transformers invertible 1x1 convolutions permutation-invariant transformers L2 perturbation PAC-Bayesian Certificates for Quadratic Closed-Loop Control Force-Informed Re-Sampling Training Training-Free Semantic Correction for Autoregressive Visual Models Intervention-Aware Variational Quantum Differentiable Predictive Control Graph Convolutional Attention

Recent events (1)

4arXiv · cs.LG·Jul 21, 2026·source ↗

Certified Training for Convolutional Perturbations achieves 80%+ robust accuracy against motion blur on CIFAR10

A new arXiv preprint introduces a certified training method that encodes convolutional perturbations (e.g., motion blur) to train vision models with formal robustness guarantees. The approach outperforms adversarial training, achieving over 80% robust accuracy against motion blur on CIFAR10 while preserving standard accuracy. The work targets safety-critical deployment scenarios where empirical robustness methods like data augmentation are insufficient due to lack of formal guarantees.

Evaluation and Benchmarking AI Safety Research Certified Training for Convolutional Perturbations CIFAR10