Abstract: Software defect prediction is an important method in the field of software security. Most of existing defect prediction models are source-oriented and can not be easily used for Android binary files (apks) defect prediction. Moreover, the traditional machine learning techniques used in these models have a shallow architecture, which leads to a limited capacity of expressing complex functions between features and defects. The author proposes a practical defect prediction model for Android binary files using deep neural network (DNN). A new approach is proposed to generate features that capture both token and semantic features of the defective smali (decompiled files of apks) files in apks. The feature vectors are input into DNN to train and build the defect prediction model in order to achieve accuracy. The article implements the model called DefectDroid and applies it to a large number of Android smali files. The performance of DefectDroid is compared from three aspects:within-project defect prediction, cross-project defect prediction and traditional machine learning algorithms.

Key words: defect prediction, software security, Android binary files, machine learning, deep neural network