What Can the Neural Tangent Kernel Tell Us About Adversarial Robustness?

Authors: Nikolaos Tsilivis, Julia Kempe

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments confirm a very strong alignment of loss gradients from the neural nets and the NTK across the whole duration of training, as can be seen in Fig. 3 (top). Then, as expected, kernel-generated attacks produce a similar drop in accuracy throughout training as the networks own white-box attacks, eventually driving robust accuracy to 0%, as seen in Fig. 3 (bottom).
Researcher Affiliation Academia Nikolaos Tsilivis Center for Data Science New York University nt2231@nyu.edu Julia Kempe Center for Data Science and Courant Institute of Mathematical Sciences New York University kempe@nyu.edu
Pseudocode No No pseudocode or algorithm blocks are present in the paper.
Open Source Code Yes We provide code to visualize features induced by kernels, giving a unique and principled way to inspect features induced by standardly trained nets (available at https://github.com/ Tsili42/adv-ntk).
Open Datasets Yes To this end, we consider classification problems from MNIST (10 classes) and CIFAR-10 (car vs airplane). We compose the Gram matrices from the whole training dataset (50000 and 10000, respectively)...
Dataset Splits No The paper mentions using 'a hold-out validation set' but does not provide specific details on the dataset splits (percentages, sample counts, or citations to predefined splits).
Hardware Specification No The paper states support from 'NYU IT High Performance Computing resources', but does not specify any particular GPU models, CPU models, or other detailed hardware specifications used for the experiments.
Software Dependencies No The paper mentions the use of 'Neural Tangents library' and 'JAX' but does not provide specific version numbers for these software dependencies.
Experiment Setup No The paper mentions 'small learning rate' and 'PGD with an ℓ∞ constraint' for training but defers 'Full implementations details' to Appendix E, which is not provided, thus lacking specific hyperparameter values in the main text.