Leveraging Satisfiability Modulo Theory Solvers for Verification of Neural Networks in Predictive Maintenance Applications

Abstract

Interest in machine learning and neural networks has increased significantly in recent years. However, their applications are limited in safety-critical domains due to the lack of formal guarantees on their reliability and behavior. This paper shows recent advances in satisfiability modulo theory solvers used in the context of the verification of neural networks with piece-wise linear and transcendental activation functions. An experimental analysis is conducted using neural networks trained on a real-world predictive maintenance dataset. This study contributes to the research on enhancing the safety and reliability of neural networks through formal verification, enabling their deployment in safety-critical domains.

NeVer2: Learning and Verification of Neural Networks

“…As witnessed by an extensive recent survey by Huang et al (2020) of more than 200 papers, the response from the scientific community to the problem of ensuring correct behavior of DNNs has been substantial. Verification (Bak et al, 2020, Demarchi et al, 2022, Eramo et al, 2022, Ferrari et al, 2022, Guidotti, 2022, Guidotti et al, 2019b,b, 2020, 2023c,d,e, Henriksen and Lomuscio, 2021, Katz et al, 2019, Kouvaros et al, 2021, Singh et al, 2019a, which aims to provide formal assurances regarding the behavior of neural networks, has emerged as a potential solution to the aforementioned robustness issues. In addition to the development of verification tools and techniques, a substantial amount of research is also directed towards modifying networks to align with specified criteria (Guidotti et al, 2019a,b, Henriksen et al, 2022, Kouvaros et al, 2021, Sotoudeh and Thakur, 2021, and exploring methods for training networks that adhere to specific constraints on their behavior (Cohen et al, 2019, Eaton-Rosen et al, 2018, Giunchiglia and Lukasiewicz, 2021, Giunchiglia et al, 2022, Hu et al, 2016.…”