Adversarially Robust Generalization Requires More Data

Authors: Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, Aleksander Madry

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We complement our theoretical results with experiments on popular image classification datasets and show that a similar gap exists here as well. To complement our theoretical results, we conduct a range of experiments on MNIST, CIFAR10, and SVHN.
Researcher Affiliation Collaboration Ludwig Schmidt UC Berkeley ludwig@berkeley.edu Shibani Santurkar MIT shibani@mit.edu Dimitris Tsipras MIT tsipras@mit.edu Kunal Talwar Google Brain kunal@google.com Aleksander M adry MIT madry@mit.edu
Pseudocode No The paper contains mathematical definitions, theorems, and experimental descriptions, but no structured pseudocode or algorithm blocks are present.
Open Source Code No The paper does not contain any statement about releasing source code for its methodology, nor does it provide any links to a code repository.
Open Datasets Yes We consider standard convolutional neural networks and train models on datasets of varying complexity. Specifically, we study the MNIST [34], CIFAR-10 [33], and SVHN [40] datasets.
Dataset Splits No The paper mentions generating training subsets by 'randomly sub-sampling the complete dataset' and evaluating performance on 'test accuracy', but it does not provide specific percentages or counts for training, validation, or test splits.
Hardware Specification No The paper does not provide specific details about the hardware used for running experiments, such as GPU or CPU models, or memory specifications.
Software Dependencies No The paper does not provide specific software dependency details with version numbers (e.g., library names with versions).
Experiment Setup Yes We perform robust optimization to train our classifiers on perturbations generated by projected gradient descent. [...] For each choice of training set size N and fixed attack εtest, we select the best performance achieved across all hyperparameters settings (training perturbations εtrain and model size).