Challenging SMT solvers to verify neural networks

In this paper we evaluate state-of-the-art SMT solvers on encodings of verification problems involving Multi-Layer Perceptrons (MLPs), a widely used type of neural network. Verification is a key technology to foster adoption of MLPs in safety-related applications, where stringent requirements about performance and robustness must be ensured and demonstrated. In previous contributions, we have shown that safety problems for MLPs can be attacked by solving Boolean combinations of linear arithmetic constraints. However, the generated encodings are hard for current state-of-the-art SMT solvers, limiting our ability to verify MLPs in practice. The experimental results herewith presented are meant to provide the community with a precise picture of current achievements and standing open challenges in this intriguing application domain.