Adversarial robustness via robust low rank representations

Authors: Pranjal Awasthi, Himanshu Jain, Ankit Singh Rawat, Aravindan Vijayaraghavan

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically demonstrate the improvements obtained by our approach on image data in Section 2. Empirical Evaluation. We compare Algorithm 1 with the algorithm of [12] for various values of σ and ε (used for training to optimize (7)). We train a Res Net-32 network on the CIFAR-10 dataset by optimizing (7). In Figure 2 we present the result of our training procedure for various values of ε and σ and compare with the ℓ2 smoothing method of [12] on the CIFAR-10 and CIFAR-100 datasets. We evaluate our approach on the CIFAR-10 and CIFAR-100 datasets.
Researcher Affiliation Collaboration Pranjal Awasthi Google Research and Rutgers University. Himanshu Jain Google Research. Ankit Singh Rawat Google Research. Aravindan Vijayaraghavan Northwestern University.
Pseudocode Yes Algorithm 1 Adversarial training via projections. Algorithm 2 Fast Certification of ℓ1 norm and Quadratic Programming.
Open Source Code No No explicit statement or link providing access to the open-source code for the methodology described in this paper.
Open Datasets Yes We train a Res Net-32 network on the CIFAR-10 dataset by optimizing (7). In Figure 2 we present the result of our training procedure for various values of ε and σ and compare with the ℓ2 smoothing method of [12] on the CIFAR-10 and CIFAR-100 datasets.
Dataset Splits No No explicit details on train/validation/test dataset splits (e.g., percentages, sample counts, or specific split files) are provided in the paper.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) used for running experiments are provided in the paper.
Software Dependencies No No specific software dependencies with version numbers (e.g., 'PyTorch 1.9', 'Python 3.8') are provided in the paper.
Experiment Setup Yes We compare Algorithm 1 with the algorithm of [12] for various values of σ and ε (used for training to optimize (7)). See Appendix B for a description of the hyperparameters and additional experiments. Following [6, 12], the inner maximization of finding adversarial perturbations is solved via projected gradient descent (PGD), and given the adversarial perturbations, the outer minimization uses stochastic gradient descent.