VNN: Verification-Friendly Neural Networks with Hard Robustness Guarantees

Authors: Anahita Baninajjar, Ahmed Rezine, Amir Aminifar

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we assess the performance of our proposed framework to generate VNNs compared to state-of-the-art techniques. Our framework is implemented using the Gurobi solver (Gurobi Optimization, LLC, 2023) on a Mac Book Pro with an 8-core CPU and 32 GB of RAM1. [...] We consider three public datasets for evaluation of VNNs, namely, the MNIST dataset (Le Cun, 1998) and two datasets from safety-critical medical applications.
Researcher Affiliation Academia 1Department of Electrical and Information Technology, Lund University, Lund, Sweden, 2Department of Computer and Information Science, Link oping University, Link oping, Sweden.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes The code is available on https://github.com/ anahitabn94/VNN.
Open Datasets Yes We consider three public datasets for evaluation of VNNs, namely, the MNIST dataset (Le Cun, 1998) and two datasets from safety-critical medical applications. The first one is about epileptic seizure detection based on the CHB-MIT Scalp EEG database (Shoeb, 2010) and the second one concerns cardiac arrhythmia detection based on the MITBIH Arrhythmia database (Goldberger et al., 2000).
Dataset Splits Yes For each patient, we consider 60%, 20%, and 20% of the dataset for the training, validation, and test sets, respectively. [...] The training set includes 75% of the entire dataset. The remaining 25% are equally partitioned among the test set and the validation set.
Hardware Specification Yes Our framework is implemented using the Gurobi solver (Gurobi Optimization, LLC, 2023) on a Mac Book Pro with an 8-core CPU and 32 GB of RAM1.
Software Dependencies Yes Our framework is implemented using the Gurobi solver (Gurobi Optimization, LLC, 2023) on a Mac Book Pro with an 8-core CPU and 32 GB of RAM1.
Experiment Setup Yes Here, we investigate how changing the hyperparameter ϵ impacts verification-friendliness on FNNs characterized by different structures. [...] For example, when ϵ = 0.1, the value of each neuron x(l) i can change in the range of [0.9x(l) i , 1.1x(l) i ]...