SOFTWARE DEFECT PREDICTION USING DEEP NEURAL NETWORKS

Abstract

Software life cycle is a long series of steps performed to guarantee a reliable, correct and a robust software. Teams of developers and quality assurance specialist are working towards a common goal of providing quality assured software. Traditionally over a period of time quality assurance has drastically improved, this involved not only focusing on what is being developed but also on how it is being developed. With ever increasing demands to deliver a quality-based customer centric product it is essential to devise new ways to achieve greater quality in less time or we say on time. Today's software systems are made up of large subsystems interlinked together to achieve a common goal. Once devised developers need to spend a huge amount of time in only finding the location of a defect. A defect if goes unnoticed can and will cause organizations to spend not only considerable amount of time and money to drill down to the root cause causing huge delays. Even if the documentations, functions or code snippets are reviewed from early stages there is always a chance that a defect goes unnoticed in development phase. Thus, this is even more important for quality assurance teams to predict the nature of change of software components in time. With keeping this is mind and to improve software reliability, various defect prediction techniques has been utilized over time by developers and quality assurance teams to assist in finding defects and properly channeling the testing efforts. With the help of these software metrics data from Statistical analysis, software change prediction model can be generated that can be useful in predicting issues in later releases of same software. Thus, the development of predictive models to predict faulty or defective classes can help & guide the stakeholders in early phase of the software development cycle. The objective of thesis is to do statistical analysis of Android data sets that are generated on Android applications like Bluetooth, Contacts, Gallery, Messaging, Music and Settings. Analyzing code and binary together we will build Deep Learning (DL) based models and fine tune parameters to better understand effects of DL techniques over Defect prediction. The evaluation is performed with an intention to find the effectiveness of the DL based model for prediction of classes’ change in software based on software quality metrics. Software quality metrics used in our study are CKJM, McCabe, Halstead [1] on Android Oreo (8.1) and Android Pie (9.0) releases.

Deep learning model was generated by first generating metrics on Android data sets and then building the DL models using Deeplearning4J library. Various models then compared together for performance.