Benchmarking provable resilience in convolutional neural networks: A study with Beta-CROWN and ERAN

Santosh Appachu Devanira Poovaiah *

University of Southern California, Los Angeles, California, USA.
 
Review
International Journal of Science and Research Archive, 2022, 07(02), 834-845.
Article DOI: 10.30574/ijsra.2022.7.2.0319
Publication history: 
Received on 11 September 2022; revised on 20 November 2022; accepted on 26 November 2022
 
Abstract: 
Robust verification of neural networks is a critical property, particularly in the context of safety-critical applications. Despite its importance, achieving both accurate and efficient verification remains a significant challenge. This study focuses on the robustness of convolutional neural networks (CNNs), which are widely deployed in domains such as autonomous driving and medical diagnostics, where failure is not an option. We explored two state-of-the-art verification techniques, namely Beta-CROWN and ERAN, that leverage linear approximation methods to handle the inherent non-linearity of neural networks. Both approaches approximate nonlinear constraints to enable tractable formal verification. A comparative analysis is conducted to evaluate their performance in terms of certification accuracy and computational efficiency. Our findings provide insights into the practical trade-offs between these techniques and guide future work in provable resilience of CNNs.
 
Keywords: 
Convolutional neural networks; Certified robustness; Formal verification; Perturbations; Rectified Linear Unit (ReLU); Beta-CROWN; ERAN
 
Full text article in PDF: 
 
Copyright information: 

Copyright © 2022 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution Liscense 4.0