Abstract Interpretation Workshop - Introduction to Neural Network Verification
Workshop Talk, University of Luxembourg, Belval, Luxembourg
Neural networks have become powerful and popular machine learning techniques in recent years. There are many safety-critical applications based on neural networks, such as autonomous driving systems. However, they can be vulnerable to adversarial examples, which are instances with imperceptible perturbations that can lead to incorrect results from neural networks. Therefore, in these safety applications, if any adversarial examples exist, the neural network may make wrong decisions, potentially causing serious consequences. Nowadays, many algorithms and frameworks have been proposed to verify the robustness of neural networks. In this talk, we will introduce how to use abstract interpretation to evaluate the robustness of neural networks. Additionally, we will provide a simple example to demonstrate how to construct a zonotope, which is one of the abstract domains, to verify neural networks.