Consistency Regularization for Certified Robustness of Smoothed Classifiers

Authors: Jongheon Jeong, Jinwoo Shin

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments under various deep neural network architectures and datasets show that the certified ℓ2-robustness can be dramatically improved with the proposed regularization, even achieving better or comparable results to the state-of-the-art approaches with significantly less training costs and hyperparameters.
Researcher Affiliation Academia School of Electrical Engineering Graduate School of AI Korea Advanced Institute of Science and Technology (KAIST) Daejeon, South Korea
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks (clearly labeled algorithm sections or code-like formatted procedures).
Open Source Code Yes Code is available at https://github.com/jh-jeong/smoothing-consistency.
Open Datasets Yes We verify the effectiveness of our proposed regularization based on extensive evaluation covering MNIST [21], CIFAR-10 [20], and Image Net [30] classification datasets.
Dataset Splits No The paper mentions training and testing on datasets like CIFAR-10 and ImageNet, and refers to following training details from prior works [10, 32], but does not explicitly provide specific train/validation/test dataset splits within its own text.
Hardware Specification Yes In this experiment, every model is trained on CIFAR-10 using one GPU of NVIDIA TITAN X (Pascal).
Software Dependencies No The paper mentions using well-known models like ResNet but does not provide specific version numbers for software dependencies or libraries used in the experiments.
Experiment Setup Yes For a fair comparison, we follow the same training details used in Cohen et al. [10] and Salman et al. [32]. For each model configuration, we consider three different models as varying the noise level σ {0.25, 0.5, 1.0}. During inference, we apply randomized smoothing with the same σ used in the training. When our regularization is used, we use m = 2 and η = 0.5 unless otherwise specified.