Adversarial Neural Pruning with Latent Vulnerability Suppression

Authors: Divyam Madaan, Jinwoo Shin, Sung Ju Hwang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our Adversarial Neural Pruning with Vulnerability Suppression (ANP-VS) method on multiple benchmark datasets, on which it not only obtains state-of-the-art adversarial robustness but also improves the performance on clean examples, using only a fraction of the parameters used by the full network.
Researcher Affiliation Collaboration 1School of Computing, KAIST, South Korea 2School of Electrical Engineering, KAIST, South Korea 3Graduate School of AI, KAIST, South Korea 4AITRICS, South Korea.
Pseudocode Yes Algorithm 1 Adversarial training by ANP-VS
Open Source Code Yes the code is available online 2. 2https://github.com/divyam3897/ANP_VS
Open Datasets Yes 1. MNIST. This dataset (Le Cun, 1998)... 2. CIFAR-10. This dataset (Krizhevsky, 2012)... 3. CIFAR-100. This dataset (Krizhevsky, 2012)
Dataset Splits No The paper mentions 'training instances' and 'test instances' for the datasets, but does not explicitly provide details for a separate validation split used in their experiments.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No The paper mentions software components like 'Le Net 5-Caffe', 'PyTorch', and 'PGD' but does not specify their version numbers.
Experiment Setup Yes For MNIST, we consider a Lenet-5-Caffe model with a perturbation radius of ε = 0.3, perturbation per step of 0.01, 20 PGD steps for training, and 40 PGD steps with random restarts for evaluating the trained model. ... We use ε = 0.03, 10 PGD steps for training and 40 steps with random restart for evaluation.