Model Checking Deep Neural Networks: Opportunities and Challenges

Zohra Sbaï ^1,2*

• ¹ Department of Computer Science, Prince Sattam Bin Abdulaziz University, Al-Kharj, Saudi Arabia
• ² Tunis El Manar University, Tunis, Tunisia

Deep neural networks (DNN) are extensively used in both current and future manufacturing, transportation, and health care sectors. The current widespread use of neural networks in highly safety-critical applications has made it necessary to prevent catastrophic issues from arising during prediction processes. In fact, misreading a traffic sign by an autonomous car or performing an incorrect analysis of medical records could put human lives in danger. Being aware of this, the number of studies related to deep neural network verification has increased dramatically in recent years. In particular, formal guarantees regarding the behavior of a DNN under particular settings are provided by model checking, which is crucial in applications that are safety-critical and where network output errors could have disastrous effects. Model checking is an effective approach for confirming that neural networks perform as planned by comparing them to clearly stated qualities. This paper aims to highlight the critical need and present challenges of using model checking verification techniques to verify deep neural networks before relying on them in real-world applications. It examines the state-of-the-art researches and draws the most prominent future directions in model checking of neural networks.