ABSTRACT. Since a few years neural networks are the most powerful tool for perception tasks, especially in image processing, and superior performances in these tasks sparked the desire to use them in safety-critical systems, e.g., autonomous vehicles. However, verifying the safety of systems that are using neural networks remains a challenge, because neural networks raise certain dependability concerns (such as adversarial inputs). Resulting from this need, the research topic of formal verification of neural networks emerged. We identify some of the main challenges of this field and discuss how to address them.