VERONA: A Framework for Neural Network Robustness Experiments
Aaron Berger, Annelot Willemijn Bosman, Holger Hoos, Jan N van Rijn
Abstract
This proposal suggests a presentation of VERONA, a framework designed to simplify robustness evaluation. In this presentation, we will describe the types of experiments for which VERONA can be used and highlight its potential for various robustness evaluation experiments by using verification, certification and adversarial attack methods. Further, in this presentation, we will provide an overview of our open-source, object-oriented software package from both a developer’s and a user’s perspective. We do this by providing a live-coding tutorial of how to use the software for verification experiments (if the alloted time permits it) and explaining how the community can contribute to the software. Additionally, given enough time, we will showcase several case studies to demonstrate the versatility and effectiveness of VERONA in different robustness evaluation experiments. In this proposal we give a motivation for using the software and an outline for the presentation.
The proposal for this presentation is based on the software that was developed for performing experiments and used (amongst others) in a paper that is under review at the Journal of Artificial Intelligence Research (JAIR). Currently, the software is also prepaired for submission to the Machine Learning Open Source Software (MLOSS) journal.
We think presenting the software at the SAIV symposium can give us insights in making it more accesible and widely-used. We are hoping for valuable feedback from the community which we can use to extend the functionalities of the software, which can help us improve the usability and community adaptation. The software has helped us set-up experiments for verification in matters of minutes instead of the multiple days or even weeks that preparing experiments used to take us. We expect that the SAIV community can also benefit from using our software and hope to show this during the presentation.