Stochastic Activation Pruning for Robust Adversarial Defense

Authors: Guneet S. Dhillon, Kamyar Azizzadenesheli, Zachary C. Lipton, Jeremy D. Bernstein, Jean Kossaifi, Aran Khanna, Animashree Anandkumar

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments to evaluate SAP address two tasks: image classification and reinforcement learning. We apply the method to the Re LU activation maps at each layer of the pretrained neural networks. To create adversarial examples in our evaluation, we use FGSM, x = λ sign(J (Mp(θ), x, y)).
Researcher Affiliation Collaboration Guneet S. Dhillon1,2, Kamyar Azizzadenesheli3, Zachary C. Lipton1,4, Jeremy Bernstein1,5, Jean Kossaifi1,6, Aran Khanna1, Anima Anandkumar1,5 1Amazon AI, 2UT Austin, 3UC Irvine, 4CMU, 5Caltech, 6Imperial College London
Pseudocode Yes Algorithm 1 Stochastic Activation Pruning (SAP)
Open Source Code Yes All the implementations were coded in MXNet framework (Chen et al., 2015) and sample code is available at https://github.com/Guneet-Dhillon/Stochastic-Activation-Pruning
Open Datasets Yes The CIFAR-10 dataset (Krizhevsky & Hinton, 2009) was used for the image classification domain.
Dataset Splits Yes It was trained on a dataset consisting of 80% un-perturbed data and 20% adversarially perturbed data, generated on the model from the previous epoch, with λ = 2. This achieved an accuracy of 75.0% on the un-perturbed validation set.
Hardware Specification No The paper states, 'Computing a single backward pass of the SAP-100 model for 512 examples takes 20 seconds on 8 GPUs.' While GPUs are mentioned, no specific model numbers, memory details, or other hardware specifications are provided.
Software Dependencies No The paper mentions 'MXNet framework (Chen et al., 2015)' but does not provide a specific version number for this framework or any other software dependencies.
Experiment Setup Yes We trained a Res Net-20 model (He et al., 2016) using SGD, with minibatches of size 512, momentum of 0.9, weight decay of 0.0001, and a learning rate of 0.5 for the first 100 epochs, then 0.05 for the next 30 epochs, and then 0.005 for the next 20 epochs.