Certified Adversarial Robustness Under the Bounded Support Set

Authors: Yiwen Kou, Qinyuan Zheng, Yisen Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We present experimental results on CIFAR-10 dataset with Res Net model to validate part of our theory about uniform smoothing measures with l2 ball and l ball support set on l2 adversary and use Gaussian smoothing measure as contrast.
Researcher Affiliation Academia 1Yuanpei College, Peking University 2Key Lab. of Machine Perception (Mo E), School of Artificial Intelligence, Peking University. 3Institute for Artificial Intelligence, Peking University.
Pseudocode Yes Algorithm 1 Certification Process
Open Source Code No The paper mentions using an implementation from GitHub for comparison purposes but does not state that their own methodology's code is open-source or provided.
Open Datasets Yes We choose CIFAR-10 as our main dataset and Res Net-110 as our base classifier.
Dataset Splits No The paper states: 'We first train the base classifier on the 50000 image training set without smoothing and achieve 89.6% prediction accuracy on the 10000 image test set.' It specifies training and test sets but does not explicitly mention a separate validation set or its split details.
Hardware Specification Yes All training, testing, and certification are run on an NVIDIA RTX 3090.
Software Dependencies No The paper does not list specific software dependencies with version numbers.
Experiment Setup Yes We set the sample amount n to 100, 1000, and 10000 with three different smoothing distributions, and they all obtain similar results: it takes only 10 minutes to run through the 10000 images test set with 100 samples for each image, 30 minutes with 1000 samples and 3 hours with excessive 10000 samples. We first implement our framework with Gaussian smoothing measure N(x, σ2I) where σ = 0.025, 0.05, 0.1 and sample amount n=100. Next, for smoothing process, we substitute Gaussian distribution with l2, l norm ball support set uniform distribution, with r = 0.025, 0.05, 0.1.