Certifiably Adversarially Robust Detection of Out-of-Distribution Data

Authors: Julian Bitterwolf, Alexander Meinke, Matthias Hein

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

Reproducibility Variable Result LLM Response
Research Type Experimental We provide experimental results for image recognition tasks with MNIST [22], SVHN [30] and CIFAR-10 [21] as in-distribution datasets.
Researcher Affiliation Academia Julian Bitterwolf University of Tübingen Alexander Meinke University of Tübingen Matthias Hein University of Tübingen
Pseudocode No The paper contains mathematical equations and descriptions of the method but no explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code is available under https://gitlab.com/Bitterwolf/GOOD.
Open Datasets Yes We provide experimental results for image recognition tasks with MNIST [22], SVHN [30] and CIFAR-10 [21] as in-distribution datasets. ... For the training out-distribution, we use 80 Million Tiny Images (80M) [37]. ... As OOD evaluation sets we use Fashion MNIST [39], the Letters of EMNIST [5], grayscale CIFAR-10, and Uniform Noise for MNIST, and CIFAR-100 [21], CIFAR-10/SVHN, LSUN Classroom [40], and Uniform Noise for SVHN/CIFAR-10.
Dataset Splits No The paper mentions training and test sets but does not explicitly describe a separate validation dataset split or how it was used for hyperparameter tuning. Standard splits are implied for the well-known datasets, but no specific validation split details are provided.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, or cloud computing instance specifications used for running the experiments.
Software Dependencies No The paper mentions software components like ReLU, softmax, Adam, Auto-PGD, and scikit-learn, but does not specify version numbers for any of them.
Experiment Setup Yes For MNIST, we use the large architecture from [13], and for SVHN and CIFAR-10 a similar but deeper and wider model. The layer structure is laid out in Table 2 in the appendix. Data augmentation is applied to both inand out-distribution images during training. For MNIST we use random crops to size 28 28 with padding 4 and for SVHN and CIFAR-10 random crops with padding 4 as well as the quite aggressive augmentation Auto Augment [9]. Additionally, we apply random horizontal flips for CIFAR-10. ... we use linear ramp up schedules for ϵ and κ, which are detailed in Appendix D. As radii for the l -perturbation model on the out-distribution we use ϵ = 0.3 for MNIST, ϵ = 0.03 for SVHN and ϵ = 0.01 for CIFAR-10. As parameter κ for the trade-off between cross-entropy loss and the GOOD regularizer in (9) and (10), we set κ = 0.3 for MNIST and κ = 1 for SVHN and CIFAR-10. A training batch consists of 128 inand 128 out-distribution samples.