Certified Training for Convolutional Perturbations
The paper claims formal robustness gains against blur-like runtime distortions without giving up standard accuracy.
Brückner and Lomuscio introduce certified training for convolutional perturbations, framed around failures such as motion blur from a shaking camera. The method uses an efficient encoding of those perturbations to train models with provable robustness guarantees. In the abstract’s CIFAR10 example, it reports more than 80% robust accuracy against reasonably intense motion blur while keeping standard accuracy comparable. ArXiv · AI/CL/LG's note
Brückner and Lomuscio introduce certified training for convolutional perturbations, framed around failures such as motion blur from a shaking camera. The method uses an efficient encoding of those perturbations to train models with provable robustness guarantees. In the abstract’s CIFAR10 example, it reports more than 80% robust accuracy against reasonably intense motion blur while keeping standard accuracy comparable. ArXiv · AI/CL/LG's note
score 5