Efficient Exact Verification of Binarized Neural Networks

Authors: Kai Jia, Martin Rinard

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of EEV by presenting the first exact verification results for ℓ -bounded adversarial robustness of nontrivial convolutional BNNs on the MNIST and CIFAR10 datasets. Our experimental results show that, compared to exact verification of robustly trained real-valued networks with the same architectures, EEV delivers several orders of magnitude faster verification of BNNs with comparable verifiable accuracy in most cases.
Researcher Affiliation Academia Kai Jia MIT CSAIL 32 Vassar St, Cambridge, MA 02139 jiakai@mit.edu Martin Rinard MIT CSAIL 32 Vassar St, Cambridge, MA 02139 rinard@csail.mit.edu
Pseudocode No The paper describes various algorithms and methods (e.g., CDCL, Bin Mask, CBD loss, adaptive gradient cancelling) but does not present any of them in a structured pseudocode block or an explicitly labeled 'Algorithm' section.
Open Source Code Yes Our source code is available at https://github.com/jia-kai/eevbnn.
Open Datasets Yes We demonstrate the effectiveness of EEV by presenting the first exact verification results for ℓ -bounded adversarial robustness of nontrivial convolutional BNNs on the MNIST and CIFAR10 datasets.
Dataset Splits Yes We evaluate the performance of EEV on the task of verifying robustness of BNNs against ℓ -bounded input perturbations on the MNIST and CIFAR10 benchmarks.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for conducting the experiments. It only discusses the theoretical aspects and software implementations.
Software Dependencies No Mini Sat CS extends Mini Sat 2.2 which is a highly optimized and minimalistic SAT solver... and ...the Glucose SAT solver [3] with the sequential counters encoding. While it mentions these, it does not provide a comprehensive list of software dependencies with specific version numbers (e.g., Python version, ML framework version, CUDA version) needed to fully replicate the environment.
Experiment Setup Yes We set τ = 5 in our experiments because it is an empirically good choice among a few values that enables fine control of the accuracy-speed tradeoff by tuning η. We present details of the experimental settings in the supplementary material.