Certifying Global Robustness for Deep Neural Networks
Published in arXiv preprint arXiv:2405.20556, 2024
Authors: You Li, Guannan Zhao, Yunqi He, Hai Zhou
Venue: arXiv preprint arXiv:2405.20556, 2024
Download: arXiv
A globally robust deep neural network resists adversarial perturbations across its entire input space, a substantially stronger guarantee than local robustness certified around individual samples. This work develops an approach to certify global robustness for deep neural networks, bringing formal-methods reasoning to bear on the reliability of learned models and complementing our broader line of work on verification and secure hardware for AI systems.