Robustness Certificates for Sparse Adversarial Attacks by Randomized Ablation

Authors: Alexander Levine, Soheil Feizi4585-4593

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally, on MNIST, we can certify the classifications of over 50% of images to be robust to any distortion of at most 8 pixels. [...] In this section, we provide experimental results of the proposed method on MNIST, CIFAR-10, and Image Net.
Researcher Affiliation Academia Alexander Levine, Soheil Feizi University of Maryland, College Park {alevine0, sfeizi}@cs.umd.edu
Pseudocode No The paper describes procedures and theoretical foundations but does not include a dedicated pseudocode block or algorithm listing.
Open Source Code Yes Code and supplementary material is available at https://github.com/alevine0/randomizedAblation/.
Open Datasets Yes Experimentally, on MNIST, we can certify the classifications... In this section, we provide experimental results of the proposed method on MNIST, CIFAR-10, and Image Net.
Dataset Splits No The paper mentions using a 'test set' for MNIST/CIFAR and the 'ILSVRC2012 validation set' for testing ImageNet, but it does not provide explicit details about standard training/validation/test splits (e.g., percentages or sample counts for validation data) for all datasets needed for reproduction.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper mentions using ResNet architectures and the Foolbox package, but it does not provide specific version numbers for any software dependencies required to reproduce the experiments.
Experiment Setup Yes Unless otherwise stated, the uncertainty α is 0.05, and 10,000 randomly-ablated samples are used to make each prediction. [...] We use 1,000 and 10,000 samples, respectively, for these two steps. [...] For performance reasons, during training, we ablate the same pixels from all images in a minibatch. We use the same retention constant k during training as at test time. [...] for greyscale images where pixels in S are floating point values between zero and one (i.e. S = [0, 1]), we encode s S as the tuple (s, 1 s), and then encode NULL as (0, 0).