Ranking source code static analysis warnings for continuous monitoring of free/libre/open source software repositories

• Ano de defesa: 2018
• Autor(a) principal: Ribeiro, Athos Coimbra
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Análise estática de código-fonte; Engenharia de software; Qualidade de software; Software engineering; Software quality; Source code static analysis
• Link de acesso: http://www.teses.usp.br/teses/disponiveis/45/45134/tde-20082018-170140/

While there is a wide variety of both open source and proprietary source code static analyzers available in the market, each of them usually performs better in a small set of problems, making it hard to choose one single tool to rely on when examining a program. Combining the analysis of different tools may reduce the number of false negatives, but yields a corresponding increase in the number of false positives (which is already high for many tools). An interesting solution, then, is to filter these results to identify the issues least likely to be false positives. This work presents kiskadee, a system to support the usage of static analysis during software development by providing carefully ranked static analysis reports. First, it runs multiple static analyzers on the source code. Then, using a classification model, the potential bugs detected by the static analyzers are ranked based on their importance, with critical flaws ranked first, and potential false positives ranked last. To train kiskadee\'s classification model, we post-analyze the reports generated by three tools on synthetic test cases provided by the US National Institute of Standards and Technology. To make our technique as general as possible, we limit our data to the reports themselves, excluding other information such as change histories or code metrics. The features extracted from these reports are used to train a set of decision trees using AdaBoost to create a stronger classifier, achieving 0.8 classification accuracy (the combined false positive rate from the used tools was 0.61). Finally, we use this classifier to rank static analyzer alarms based on the probability of a given alarm being an actual bug. Our experimental results show that, on average, when inspecting warnings ranked by kiskadee, one hits 5.2 times less false positives before each bug than when using a randomly sorted warning list.