Certified Robustness for Deep Equilibrium Models via Interval Bound Propagation

Authors: Colin Wei, J Zico Kolter

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical comparison reveals that models with IBP-Mon DEQ layers can achieve comparable ℓ8 certified robustness to similarly-sized fully explicit networks.1 Our experiments demonstrate that IBP-Mon DEQ layers are competitive with standard explicit layers for ℓ8-certified robustness.
Researcher Affiliation Collaboration Colin Wei Stanford University colinwei@stanford.edu J. Zico Kolter CMU and Bosch Center for AI zkolter@cs.cmu.edu
Pseudocode No The paper does not contain any clearly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes Code is available here: https://github.com/cwein3/ibp-mondeq-code.
Open Datasets Yes Tables 2 shows certified and standard classification errors of 3 and 7 layer models trained with IBP on the MNIST and CIFAR10 datasets for various values of ϵ.
Dataset Splits Yes For CIFAR10, we choose ϵtrain ϵtest, whereas for MNIST we use a larger value of ϵtrain, following Gowal et al. (2018) and Shi et al. (2021). The values are displayed in Table 3.
Hardware Specification Yes All models besides the DEQ-3 can be trained within a day on a single NVIDIA Titan Xp GPU.
Software Dependencies No The paper mentions 'Adam optimizer (Kingma & Ba, 2014)' but does not provide specific version numbers for software dependencies such as Python, PyTorch, or other libraries.
Experiment Setup Yes We train using IBP with the Adam optimizer (Kingma & Ba, 2014) with a learning rate of 5e-4, and report errors at the last epoch of training averaged over 3 runs. ... We use a batch size of 128 and anneal the learning rate, which is initially 5e-4, by a factor of 0.2 at certain steps. ... We use gradient clipping with a max ℓ2 norm of 10.