Test Case Generation for Convolutional Neural Network

Abstract

In this paper, we present a test image generation approach for Convolutional Neural Network (CNN) that is widely used for image recognition. The goal of our approach has to generate an image satisfying the following two conditions. First, this image activates a specific cell on CNN for checking the effect of this cell. Second, this image looks like a real image as a plausible test case. For this purpose, we combine the activation maximization technique with GAN (generative adversarial network). Even though quite a large number of sample data are used for training, it is infeasible to activate every cell and check its effect. Thus, this technique is useful for verifying cells uncovered in training and, thus, for improving the quality of CNN. To our best knowledge, this is the first attempt to improve the quality of images using the generative modeling approach. With the famous MNIST example, we illustrate the details and benefits of the proposed approach.