Generative Model-Based Test Case Generation and Operational Testing for Deep Learning

As deep learning technologies, particularly those driven by deep neural networks, revolutionize various industries, testing and verifying deep learning models remain challenging, especially regarding their security. Traditional software testing methods like test case selection and generation are often unsuitable for deep learning applications due to their reliance on high-cost, manually labeled data. This paper introduces a test case generation method using deep generative models and a testing framework tailored for operational testing. This approach reduces the need for data annotation and lowers the costs associated with testing deep learning software. The framework focuses on using training data alone to assess quality metrics in operational environments, without additional annotated test cases. Experimental results confirm the method’s efficacy in accurately evaluating the performance of deep learning models and identifying regression issues, offering significant advantages over traditional testing methods.