Advanced SEU and MBU Vulnerability Assessment of Deep Neural Networks in Air-to-Air Collision Avoidance Systems via SAT-Based Techniques

Deep neural networks (DNNs) are being widely used to solve real-world challenges. Their algorithms can be impacted by environmental and cyber threats, and the difficulty of giving formal guarantees regarding their behavior under attack is a fundamental challenge for employing them in safety-critical systems. In this work, we analyze the impact of single-event upsets (SEUs) and multiple-bit upsets (MBUs) on DNNs. The verification procedure is intended to ensure that a network of interest complies with safety requirements and is dependable in critical conditions. We thus investigate the application of SAT-based analytic methodology to gain insight about the behavior and vulnerabilities of DNNs in safety-critical applications. The vertical collision avoidance system (VCAS) is used as a bench-mark to illustrate our methodology and our experimental findings show that the resilience of neural networks differ depending on the number of weight changes that occur.