Certified Adversarial Robustness with Additive Noise

Authors: Bai Li, Changyou Chen, Wenlin Wang, Lawrence Carin

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our evaluation on MNIST, CIFAR-10 and Image Net suggests that the proposed method is scalable to complicated models and large data sets, while providing competitive robustness to state-of-the-art provable defense methods. We conduct a comprehensive set of experiments to evaluate both the theoretical and empirical performance of our methods, with results that are competitive with the state of the art.
Researcher Affiliation Academia Bai Li Department of Statistical Science Duke University bai.li@duke.edu Changyou Chen Department of CSE University at Buffalo, SUNY cchangyou@gmail.com Wenlin Wang Department of ECE Duke University wenlin.wang@duke.edu Lawrence Carin Department of ECE Duke University lcarin@duke.edu
Pseudocode Yes Algorithm 1 Certified Robust Classifier
Open Source Code Yes The source code can be found at https://github.com/Bai-Li/STN-Code.
Open Datasets Yes Our evaluation on MNIST, CIFAR-10 and Image Net suggests that the proposed method is scalable to complicated models and large data sets, while providing competitive robustness to state-of-the-art provable defense methods. We perform experiments on the MNIST and CIFAR-10 data sets, to evaluate the theoretical and empirical performance of our methods. We subsequently also consider the larger Image Net dataset.
Dataset Splits No The paper mentions using a "test set" but does not specify explicit train/validation/test splits by percentages or sample counts, nor does it cite a predefined split with specific details for reproducibility beyond implicitly using standard dataset splits.
Hardware Specification No The paper does not explicitly describe the hardware used for running the experiments, such as specific GPU or CPU models.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes For the MNIST data set, the model architecture follows the models used in [36], which contains two convolutional layers, each containing 64 filters, followed with a fully connected layer of size 128. For the CIFAR-10 dataset, we use a convolutional neural network with seven convolutional layers along with Max Pooling. In both datasets, image intensities are scaled to [0, 1], and the size of attacks are also rescaled accordingly. In all our subsequent experiments, we use the end points (lower for p(1) and upper for p(2)) of the 95% confidence intervals for estimating p(1) and p(2), and multiply 95% for the corresponding accuracy. In practice, we find a sample size of n = 100 is sufficient. In previous experiments, we use σ = 0.7 and σ = 100/255 for MNIST and CIFAR-10, respectively.