Test Case Generation Method Based on TEASGD

With the development of artificial intelligence technology, the importance of DNN testing work is increasingly recognized. This article proposes a DNN testing adequacy evaluation method based on geometric diversity index. The method constructs an evaluation model through training data, which can predict the testing adequacy based on the geometric diversity of the test set. In addition, this article also proposes a test case generation method based on feature combination. Experimental results show that isolated faults can not be ignored in the test process, TEASGD algorithm is effective in the evaluation of test adequacy, and the correlation between geometric diversity and fault detection rate is more than $\mathbf{0. 7}$. Our test case generation method can improve the fault detection ability of the test set.