Certified Defense to Image Transformations via Randomized Smoothing

Authors: Marc Fischer, Maximilian Baader, Martin Vechev

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

Reproducibility Variable Result LLM Response
Research Type Experimental A thorough evaluation of all methods on common image datasets, achieving provable distributional robust accuracy of 73% for rotations with up to 30 on Restricted Image Net.
Researcher Affiliation Academia Marc Fischer, Maximilian Baader, Martin Vechev Department of Computer Science ETH Zurich {marc.fischer, mbaader, martin.vechev}@inf.ethz.ch
Pseudocode No The paper describes methods and steps in prose but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes We provide an implementation of all methods at https://github.com/eth-sri/transformation-smoothing.
Open Datasets Yes We evaluate on Image Net [30], Restricted Image Net (RImage Net)[31], a subset of Image Net with 10 classes, CIFAR-10 [32], and MNIST [33].
Dataset Splits No The paper mentions using a 'training dataset' and evaluating on 'test set' but does not provide explicit train/validation/test split percentages, absolute sample counts, or detailed methodology for dataset partitioning to ensure reproducibility of splits.
Hardware Specification Yes All experiments were performed on a machine with 2 Ge Force RTX 2080 Tis and an Intel(R) Core(TM) i9-9900K CPU.
Software Dependencies No The paper mentions using 'Py Torch [28]' and 'robustness [29]' but does not provide specific version numbers for these software components or other key dependencies.
Experiment Setup Yes Further, we let σγ, αγ, nγ, rγ and σδ, αδ, nδ, rδ denote the parameters and radius required to use Theorem 3.2 and Theorem 3.1 in practice, respectively. [...] Here we use αδ = 0.002 and αγ = 0.01 for confidences. [...] σδ = 0.3 for MNIST, σδ = 0.25 for CIFAR-10 and σδ = 0.5 for (R)Image Net. [...] For rotations (Γ = 10, σγ = 30, nγ = 2000, 3 attacks per image, 1000 images) we fix E = 0.7 and use 100 samples of β to obtain the correct αE (Eq. (9)). [...] We use 10 refinement steps and nδ = 200 for both rotations and translations.