In recent years, the ubiquity of mobile devices has witnessed a concomitant surge in malware attacks, with Android platforms emerging as predominant targets. The inherent open-source architecture of the Android operating system inadvertently paves the way for the proliferation of malevolent applications. Conventional malware detection methodologies, which predominantly hinge on manual feature extraction and juxtaposition against feature repositories, are notably resource-draining. Furthermore, an over-reliance on singular features often obfuscates the demarcation between benign and malevolent applications. To address this lacuna, this study propounds an innovative Android malicious code detection paradigm, amalgamating multifarious features with the prowess of deep learning. By decompiling APK (Android application package), we extract three static features: Opcode (operation code), API (application program interface) call, and Permission (permission). We use the N-Gram method to process the operation code to improve the extraction efficiency, filter out effective permissions to improve the classification accuracy, and select APIs related to permissions to increase the logic between features. After encoding, we input them into DNN for classification. Experiments on a dataset containing 8008 applications show that using multi-features and deep learning networks can significantly improve the accuracy of malware detection, verifying the superiority of this method in detecting malicious code.