A numerical verification method for multi-class feed-forward neural networks

• Published in: Expert systems with applications 2024-08, Vol.247, p.123345, Article 123345
• Main Authors: Grimm, Daniel, Tollner, Dávid, Kraus, David, Török, Árpád, Sax, Eric, Szalay, Zsolt
• Format: Article
• Language: English
• Subjects: Explainable neural networks; Neural network verification; Nonlinear optimization
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2024.123345

The use of neural networks in embedded systems is becoming increasingly common, but these systems often operate in safety–critical environments, where a failure or incorrect output can have serious consequences. Therefore, it is essential to verify the expected operation of neural networks before deploying them in such settings. In this publication, we present a novel approach for verifying the correctness of these networks using a nonlinear equation system under the assumption of closed-form activation functions. Our method is able to accurately predict the output of the network for given specification intervals, providing a valuable tool for ensuring the reliability and safety of neural networks in embedded systems. •A novel verification concept for neural networks is developed.•Continuous activation function based NNs can be verified.•The approach provides explainability and transparency for the verified neural network.•Monotonicity or linearity are not necessary during the verification.•No model simplification is required to evaluate the operation process.