Gradient-Guided Test Case Generation for Image Classification Software

The widespread use of deep neural networks (DNNs) in image classification sofwares underlines the importance of the robustness. Researchers have proposed sparse adversarial attack methods for generating test cases, which add pixel-level perturbations to construct the test case to mislead the target model. However, the existing methods have certain limitations, such as high time cost, poor flexibility, and poor quality of the test cases. To address these issues, we propose a gradient-guided test case generation method (GGTM) to evaluate the robustness of image classification software. The method firstly identifies the key region in the image based on the gradient-weighted class activation mapping (Grad-CAM) and the prediction confidence of the target model on the input image. In the key region, it selects a set of pixels as candidate perturbation pixels according to the gradient value and the change of loss function. Then perturbations are added to the candidate perturbation pixels after applying a random dropout strategy to reduce some candidate perturbation pixels which is used to avoid local optimum. For the initially constructed test case which can mislead the target model, after removing redundant and unimportant perturbations, perturbations are re-added to optimize the test case. Experiments show the effectiveness of GGTM, which achieves 100% attack success rate. And the test cases generated by GGTM have the best perturbation sparsity. Furthermore, compared with the baseline method SparseAG which achieves optimal perturbation sparsity among the baseline methods, GGTM significantly improves the efficiency.