Previsão de vazamento de recursos em aplicações Android usando Aprendizado de Máquina

Abstract

When mobile applications acquire device resources (such as camera, media player, and sensors) without releasing them properly and in a timely manner, a failure called resource leak occurs. This type of failure can cause serious problems, such as device performance degradation or system failure. This work proposes the LeakPred approach to assist developers in identifying components that have resource leaks. A set of six metrics related to the lifetime of resources or the application was selected to characterize the components. Six machine learning techniques were analyzed to identify leaky components from these metrics. The results suggest that the LeakPred approach, associated with classification techniques, is capable of identifying resource leaks, with two models, k-Nearest Neighbors and deep neural network, obtaining, respectively, accuracies of 87.84% and 87.75%. The LeakPred approach was compared with 5 state-of-the-art tools, namely, Android Lint, FindBugs, Infer, Checker Framework and EcoAndroid, surpassing all of them in the rate of identification of components with resource leaks.