Measuring Robustness of Deep Neural Networks from the Lens of Statistical Model Checking
Hao Bu, Meng Sun · 2023
Measuring robustness of deep neural networks (DNNs) is an important topic for trustworthy AI. Existing methods for verifying local robustness of DNNs usually face the scalability problem and have difficulties to deal with non-linear activation functions and complex semantic perturbations. Existing methods for measuring global robustness usually rely on large datasets, so it is difficult for users without a large dataset to compare the global robustness of different networks. In this paper, we propose two algorithms to measure the local and global robustness of DNNs from the lens of statistical model checking. Compared with the local robustness estimation method using the Okamoto bound, our method can provide the same theoretical guarantee using only 48.6%-74.1% of running time. Our global robustness estimation algorithm can provide high-quality estimation using only 10 images for CIFAR-10 and 50 images for ImageNet, thus can be used in a wider range of scenarios.