QuanDA: GPU Accelerated Quantitative Deep Neural Network Analysis

Over the past years, numerous studies demonstrated vulnerability of deep neural networks (DNNs) to make correct classifications in presence small noise. This motivated formal analysis DNNs ensure they delineate acceptable behavior. However, case DNN’s behavior is unacceptable for desired application, these qualitative approaches are ill-equipped determine precise degree which DNN behaves unacceptably. Towards this, we propose a novel quantitative framework, QuanDA, does not only check if delineates certain behavior, but also provides estimated probability this particular Unlike (few) available frameworks, QuanDA use any implicit assumptions on distribution hidden nodes, enables framework propagate close real distributions node values each proceeding layer. Furthermore, our leverages CUDA parallelize analysis, enabling high-speed GPU implementation fast analysis. The applicability using ACAS Xu benchmark, provide reachability estimates all network nodes. Moreover, paper potential applications safety properties.