Abstract

The analysis of Deep Neural Networks (DNNs) used in safety-critical systems using SAT techniques is covered in this thesis, which emphasizes how crucial it is to verify the networks’ safety and reliability. Ensuring the proper behavior of these networks becomes essential for the safety of humans and the integrity of the systems involved as DNNs are increasingly integrated into safety-critical systems including autonomous vehicles, medical diagnosis, and aerospace systems.
The main objective of this research is to formally verify a deep neural network, namely the Vertical Collision Avoidance System (VCAS), deployed in safety-critical applications. The verification procedure is intended to ensure that the network complies with security requirements and performs dependably under different conditions. Through the process of verification, the DNN’s predictable behavior is ensured, and the possibility of malfunctions that can result in risk or system failures is reduced. In order to get important insights into the behavior and vulnerabilities of DNNs used in safety-critical systems, this work investigates the application of SAT-based analytic methodologies.
The methodology employed in this research involves a comprehensive understanding of the DNN’s structure, training process, and inference mechanisms. By comprehending the underlying principles and potential risks associated with DNNs, the verification process can be tailored to address specific safety concerns and requirements. Furthermore, fault injection techniques, including Single Event Upsets (SEUs) and Multiple Bit Upsets (MBUs), are used to assess the network’s resilience and their ability to recover from faults, contributing to a more comprehensive understanding of DNN robustness in safety-critical contexts. This research offers invaluable insights for experts by addressing the verification difficulties of DNNs in safety-critical systems. The findings contribute to the ongoing efforts to develop standardized verification methodologies and safety guidelines for DNN deployment, ensuring the reliability, trustworthiness, and safety of these networks in critical applications.