Confidence-Aware Training of Smoothed Classifiers for Certified Robustness

Authors: Jongheon Jeong, Seojin Kim, Jinwoo Shin

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that the proposed method, despite its simplicity, consistently exhibits improved certified robustness upon state-of-the-art training methods. We evaluate the effectiveness of our proposed training scheme based on various well-established image classification benchmarks to measure robustness, including MNIST (Le Cun et al. 1998), Fashion-MNIST (Xiao, Rasul, and Vollgraf 2017), CIFAR-10/100 (Krizhevsky 2009), and Image Net (Russakovsky et al. 2015) (for certified robustness)5, as well as MNIST-C (Mu and Gilmer 2019)6 and CIFAR-10C (Hendrycks and Dietterich 2019) (for corruption robustness).
Researcher Affiliation Academia Jongheon Jeong*, Seojin Kim*, Jinwoo Shin Korea Advanced Institute of Science and Technology (KAIST) Daejeon, 34141 South Korea {jongheonj, osikjs, jinwoos}@kaist.ac.kr
Pseudocode Yes The complete procedure of computing our proposed CAT-RS loss can be found in Algorithm 1 of Appendix A.
Open Source Code Yes Code is available at https://github.com/alinlab/smoothing-catrs.
Open Datasets Yes We evaluate the effectiveness of our proposed training scheme based on various well-established image classification benchmarks to measure robustness, including MNIST (Le Cun et al. 1998), Fashion-MNIST (Xiao, Rasul, and Vollgraf 2017), CIFAR-10/100 (Krizhevsky 2009), and Image Net (Russakovsky et al. 2015) (for certified robustness)5...
Dataset Splits Yes We evaluate the performance on the uniformly-subsampled 500 samples in the Image Net validation dataset following (Cohen, Rosenfeld, and Kolter 2019; Jeong and Shin 2020; Salman et al. 2019; Jeong et al. 2021).
Hardware Specification No The paper does not explicitly mention any specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For a fair comparison, we follow the standard protocol and training setup of the previous works (Cohen, Rosenfeld, and Kolter 2019; Zhai et al. 2020; Jeong and Shin 2020).7 More details, e.g., training setups, datasets, and hyperparameters, can be found in Appendix B. We use n = 100, 000, n0 = 100, and α = 0.001 for CERTIFY, following previous works.