Condition-Guided Adversarial Generative Testing for Deep Learning Systems

Abstract

Over the past decade, Deep Neural Networks (DNNs) have achieved remarkable progress. However, the quality of such kind of systems is far from perfect. Software test is one of the most effective techniques for finding bugs in DNNs. Test case generation is the key factors of the success of software test. Existing test case generation approaches for DNNs always generate a large number of test cases, most of which do not meet the test requirements or the actual situation. In this paper, we propose CAGTest, a condition-guided adversarial generative testing tool for other DNNs to generate their test inputs to find potential defects. In general, CAGTest can generate test cases conditionally, which is not only efficient, but also does not produce a large number of invalid test cases and reduces the scale of test cases.