REGLO: Provable Neural Network Repair for Global Robustness Properties

Authors: Feisi Fu, Zhilu Wang, Weichao Zhou, Yixuan Wang, Jiameng Fan, Chao Huang, Qi Zhu, Xin Chen, Wenchao Li

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments The goal of the experiments is to validate the effectiveness of REGLO in enhancing the global robustness of trained neural networks while preserving their performance. We compare REGLO against six other methods, including four baseline methods, Certi Fair (Khedr and Shoukry 2022), a well-known method for training provably fair DNN, and VGRP (Chen et al. 2021), a training procedure to train classifiers with verified global robustness properties.
Researcher Affiliation Academia Feisi Fu1 , Zhilu Wang2 , Weichao Zhou1, Yixuan Wang2, Jiameng Fan1, Chao Huang3, Qi Zhu2, Xin Chen4, Wenchao Li1 1Department of Electrical and Computer Engineering, Boston University, Boston, MA, USA 2Department of Electrical and Computer Engineering, Northwestern University, Evanston, IL, USA 3Department of Computer Science, University of Liverpool, Liverpool, UK 4 University of New Mexico, Albuquerque, NM, USA
Pseudocode Yes Algorithm 1: Iterative Repair for Re LU DNN
Open Source Code Yes Our code is available on Git Hub: https://github.com/BUDEPEND-Lab/REGLO/tree/main/net repair.
Open Datasets Yes We train a Re LU DNN on the German Credit dataset (Dua and Graff 2017)...
Dataset Splits No The paper mentions 'training' and 'test' sets with their sizes in Table 4, but does not provide explicit train/validation/test split percentages, sample counts for validation, or references to predefined splits for reproducibility.
Hardware Specification Yes All the experiments were run on machines with CPUs similar to ten-core Intel Xeon E5-2660v3 @ 2.6 GHz with a single Ge Force GTX 1080 Graphics Card.
Software Dependencies No The paper mentions 'Python', 'Gurobi (Gurobi Optimization, LLC 2021)', and 'CVXPY (Diamond and Boyd 2016)', but does not provide specific version numbers for these software components (e.g., Python 3.x, Gurobi vX.Y, CVXPY vX.Y).
Experiment Setup Yes We consider the global robustness properties on input domain X as well as regions based on age groups... For REGLO, we search for repair areas by random sampling on X and randomly choose 30 of them to repair.