Distributionally Robust Deep Neural Network for Malicious App Software Dynamic Detection

Malicious App Software Detection (MASD) is a crucial step in ensuring device security, as it helps users identify and eliminate malicious software that may pose a threat to the system or data security. Based on the distributionally robust optimization theory, this paper constructs an ambiguity set by using Wasserstein distance, and furtherly proposes a distributionally robust Deep Neural Networks (DNN) model to deal with unseen data for the MASD problem. A computationally-efficient upper approximation of the proposed model is given based on Lipchitz constants of the DNN network. Various test results reveal that the proposed model has better performances than the corresponding DNN. And our model possesses outstanding robustness when the training data set is reduced.