A Framework of Formal Specification-Based Data Generation for Deep Neural Networks
Yanzhao Xia, Shaoying Liu · 2023
Deep Neural Networks (DNNs) have gained growing attention in many domain-specific supervised learning applications. However, the current DNNs still face two challenges. One is the difficulty of obtaining well-labeled training data for supervised learning and the other is concerned with the efficiency of training due to the lack of precise characteristics of the objects in the training process. We propose a framework of formal specification-based data generation for the training and testing of DNNs. The framework is characterized by using formal specifications to define the important and distinct features of the objects to be identified. The features are expected to serve as the foundation for generating training and testing data for DNNs. In this paper, we discuss all the activities involved in the framework and the detailed approach to writing the formal specifications. We also conduct a case study on traffic sign recognition to validate the framework.