Android mobile phones’ growing popularity has led to developers creating more malicious apps, which can be included in third-party arcades as protected applications. Detecting these malware applications is challenging due to time-consuming and high-cost techniques. This study proposes a robust deep learning (DL) model for detecting adversarial third-party apps using adaptive feature learning. The strategy involves preprocessing raw apk files, extracting permission behavioral features, and using the proposed spatial dropout-assisted convolutional autoencoder (SD_ConvAE) model to determine if the app is benign or malignant. The approach is simulated using a Python tool and assessed using various measures like accuracy, recall, weighted F-score (W-FS), false discovery rate (FDR), and kappa coefficient. The overall accuracies achieved by the developed techniques are about 99.6% and 99% for detecting benign and malignant apps, respectively.