(Certified!!) Adversarial Robustness for Free!

Authors: Nicholas Carlini, Florian Tramer, Krishnamurthy Dj Dvijotham, Leslie Rice, Mingjie Sun, J Zico Kolter

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate diffusion denoised smoothing on two standard datasets, CIFAR-10 and Image Net, and find it gives state-of-the-art certified ℓ2 robustness on both.
Researcher Affiliation Collaboration 1Google 2Carnegie Mellon University 3Bosch Center for AI
Pseudocode Yes Figure 1: Our approach can be implemented in under 15 lines of code, given an off-the-shelf classifier fclf and an off-the-shelf diffusion model denoise. The PREDICT function is adapted from Cohen et al. (2019) and takes as input a number of noise samples N and a statistical significance level η (0, 1) and inherits the same robustness certificate proved in Cohen et al. (2019).
Open Source Code Yes Code to reproduce our experiments is available at: https://github.com/ethz-privsec/diffusion_denoised_smoothing.
Open Datasets Yes We evaluate diffusion denoised smoothing on two standard datasets, CIFAR-10 and Image Net... Image Net: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248 255. Ieee, 2009.
Dataset Splits Yes Image Net configuration. We denoise Image Net images with the 552M-parameter classunconditional diffusion model from Dhariwal & Nichol (2021), and classify images with the 305Mparameter BEi T large model (Bao et al., 2022) which reaches a 88.6% top-1 validation accuracy using the implementation from timm (Wightman, 2019).
Hardware Specification Yes We obtain a throughput of 825 images per second through the diffusion model and Vi T classifier on an A100 GPU at a batch size of 1,000.
Software Dependencies No No specific version numbers for software dependencies were provided. The paper mentions: "We use the implementation from Hugging Face" and "using the implementation from timm (Wightman, 2019)".
Experiment Setup Yes On CIFAR-10, we draw N = 100,000 noise samples and on Image Net we draw N = 10,000 samples to certify the robustness following Cohen et al. (2019). ... As is standard in prior work, we perform randomized smoothing for three different noise magnitudes, σ {0.25, 0.5, 1.0}.