Denoised Smoothing: A Provable Defense for Pretrained Classifiers

Authors: Hadi Salman, Mingjie Sun, Greg Yang, Ashish Kapoor, J. Zico Kolter

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

Reproducibility Variable Result LLM Response
Research Type Experimental we demonstrate its effectiveness through extensive experimentation on Image Net and CIFAR-10. We verify the efficacy of our method through extensive experimentation on Image Net and CIFAR-10. We are able to convert pretrained Res Net-18/34/50 and Res Net-110, on CIFAR-10 and Image Net respectively, into certifiably robust models; our results are summarized in Tables 1 and 2 (details are in section 3).
Researcher Affiliation Collaboration Hadi Salman hasalman@microsoft.com Microsoft Research Mingjie Sun mingjies@cs.cmu.edu CMU Greg Yang gragyang@microsoft.com Microsoft Research Ashish Kapoor akapoor@microsoft.com Microsoft Research J. Zico Kolter zkolter@cs.cmu.edu CMU
Pseudocode No No structured pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes Our code replicating all the experiments in the paper can be found at: https: //github.com/microsoft/denoised-smoothing1.
Open Datasets Yes extensive experimentation on Image Net and CIFAR-10. We are able to convert pretrained Res Net-18/34/50 and Res Net-110, on CIFAR-10 and Image Net respectively, into certifiably robust models
Dataset Splits Yes To assess the performance of our method on these APIs, we aggregate 100 random images from the Image Net validation set and certify their predictions across all four APIs.
Hardware Specification No No specific details about GPU or CPU models, memory, or cloud instance types used for experiments were provided in the main text.
Software Dependencies No The paper mentions software like 'Py Torch-pretrained Res Net' and denoisers like 'Dn CNN' and 'Mem Net', but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes In the following experiments, we only report the results for σ = 0.25,8 and we report the best curves over the denoiser architectures mentioned above. For more details on the architectures of the classifiers/denoisers, training/certification hyperparameters, etc., we refer the reader to Appendix A.