Constructing Unrestricted Adversarial Examples with Generative Models

Authors: Yang Song, Rui Shu, Nate Kushman, Stefano Ermon

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical results on the MNIST, SVHN, and Celeb A datasets show that unrestricted adversarial examples can bypass strong adversarial training and certified defense methods designed for traditional adversarial attacks.
Researcher Affiliation Collaboration Yang Song Stanford University yangsong@cs.stanford.edu Rui Shu Stanford University ruishu@cs.stanford.edu Nate Kushman Microsoft Research nkushman@microsoft.com Stefano Ermon Stanford University ermon@cs.stanford.edu
Pseudocode Yes In what follows, we explore two attacks derived from variants of AC-GAN (see pseudocode in Appendix B).
Open Source Code No The paper does not contain any explicit statement or link indicating the availability of the source code for the described methodology.
Open Datasets Yes The datasets used in our experiments are MNIST [25], SVHN [26], and Celeb A [27].
Dataset Splits No The paper mentions using training and test partitions but does not provide specific percentages or counts for training, validation, and test splits needed for reproduction, nor does it refer to a standard split that quantifies all partitions.
Hardware Specification No The paper does not specify the exact models or types of hardware (e.g., GPUs, CPUs, memory) used for running the experiments.
Software Dependencies No The paper mentions software components like TensorFlow (Appendix C: 'We implement all models in TensorFlow.') but does not provide specific version numbers for these or any other key libraries or solvers.
Experiment Setup Yes For more details about architectures, hyperparameters and adversarial training methods, please refer to Appendix C.