Deep Neural Networks (DNNs) have been widely applied in the field of computer vision. While DNNs exhibit remarkable accuracy in classification tasks, they remain susceptible to adversarial examples. Robustness testing for DNNs has become crucial. Existing test case generation methods suffer from certain limitations, such as dispersed distribution of perturbations, insufficient realism and high time cost. This paper proposes a test case generation method SparseTG-GS (Sparse perturbation based Test case Generation - Group-wise Sparsity). Initially, superpixel segmentation is applied to partition the input image into multiple superpixels. Next, perturbation groups are iteratively introduced into the superpixels until the perturbed image effectively compromises the target DNN. Finally, based on the importance of each perturbation, the initial perturbation scheme is improved by re-adding perturbation groups and minimizing superfluous perturbations and perturbation magnitudes. We improve the AdvGAN framework to construct the Generative Adversarial Networks (GAN) for calculating a perturbation importance matrix for test cases with enhanced optimization efficiency. Experimental study shows that the test cases generated by SparseTG-GS achieve the best attack success rate for target DNNs. The perturbed superpixels in the test cases are fewer and more interpretable. In addition, SparseTG-GS is efficient for generating test cases, especially on large-sized images. The code of SparseTG-GS is available at https://github.com/Yunhelee/SparseTG-GS .
Ji et al. (Wed,) studied this question.