Fundamental Limits in Formal Verification of Message-Passing Neural Networks

Marco Sälzer, Martin Lange

Output reachability and adversarial robustness are among the most relevant safety properties of neural networks. We show that in the context of Message Passing Neural Networks (MPN…