Guiding Quality Assurance Through Context Aware Learning

Doctoral thesis (2023)

Software Testing is a quality control activity that, in addition to finding flaws or bugs, provides confidence in the software’s correctness. The quality of the developed software depends on the strength ... [more ▼]

Software Testing is a quality control activity that, in addition to finding flaws or bugs, provides confidence in the software’s correctness. The quality of the developed software depends on the strength of its test suite. Mutation Testing has shown that it effectively guides in improving the test suite’s strength. Mutation is a test adequacy criterion in which test requirements are represented by mutants. Mutants are slight syntactic modifications of the original program that aim to introduce semantic deviations (from the original program) necessitating the testers to design tests to kill these mutants, i.e., to distinguish the observable behavior between a mutant and the original program. This process of designing tests to kill a mutant is iteratively performed for the entire mutant set, which results in augmenting the test suite, hence improving its strength. Although mutation testing is empirically validated, a key issue is that its application is expensive due to the large number of low-utility mutants that it introduces. Some mutants cannot be even killed as they are functionally equivalent to the original program. To reduce the application cost, it is imperative to limit the number of mutants to those that are actually useful. Since it requires manual analysis and test executions to identify such mutants, there is a lack of an effective solution to the problem. Hence, it remains unclear how to mutate and test a code efficiently. On the other hand, with the advancement in deep learning, several works in the literature recently focused on using it on source code to automate many nontrivial tasks including bug fixing, producing code comments, code completion, and program repair. The increasing utilization of deep learning is due to a combination of factors. The first is the vast availability of data to learn from, specifically source code in open-source repositories. The second is the availability of inexpensive hardware able to efficiently run deep learning infrastructures. The third and the most compelling is its ability to automatically learn the categorization of data by learning the code context through its hidden layer architecture, making it especially proficient in identifying features. Thus, we explore the possibility of employing deep learning to identify only useful mutants, in order to achieve a good trade-off between the invested effort and test effectiveness. Hence, as our first contribution, this dissertation proposes Cerebro, a deep learning approach to statically select subsuming mutants based on the mutants’ surrounding code context. As subsuming mutants reside at the top of the subsumption hierarchy, test cases designed to only kill this minimal subset of mutants kill all the remaining mutants. Our evaluation of Cerebro demonstrates that it preserves the mutation testing benefits while limiting the application cost, i.e., reducing all cost factors such as equivalent mutants, mutant executions, and the mutants requiring analysis. Apart from improving test suite strength, mutation testing has been proven useful in inferring software specifications. Software specifications aim at describing the software’s intended behavior and can be used to distinguish correct from incorrect software behaviors. Specification inference techniques aim at inferring assertions by generating and filtering candidate assertions through dynamic test executions and mutation testing. Due to the introduction of a large number of mutants during mutation testing such techniques are also computationally expensive, hence establishing a need for the selection of mutants that fit best for assertion inference. We refer to such mutants as Assertion Inferring Mutants. In our analysis, we find that the assertion inferring mutants are significantly different from the subsuming mutants. Thus, we explored the employability of deep learning to identify Assertion Inferring Mutants. Hence, as our second contribution, this dissertation proposes Seeker, a deep learning approach to statically select Assertion Inferring Mutants. Our evaluation demonstrates that Seeker enables an assertion inference capability comparable to the full mutation analysis while significantly limiting the execution cost. In addition to testing software in general, a few works in the literature attempt to employ mutation testing to tackle security-related issues, due to the fault-based nature of the technique. These works propose mutation operators to convert non-vulnerable code to vulnerable by mimicking common security bugs. However, these pattern-based approaches have two major limitations. Firstly, the design of security-specific mutation operators is not trivial. It requires manual analysis and comprehension of the vulnerability classes. Secondly, these mutation operators can alter the program semantics in a manner that is not convincing for developers and is perceived as unrealistic, thereby hindering the usability of the method. On the other hand, with the release of powerful language models trained on large code corpus, e.g. CodeBERT, a new family of mutation testing tools has arisen with the promise to generate natural mutants. We study the extent to which the mutants produced by language models can semantically mimic the behavior of vulnerabilities aka Vulnerability-mimicking Mutants. Designed test cases failed by these mutants will also tackle mimicked vulnerabilities. In our analysis, we found that a very small subset of mutants is vulnerability-mimicking. Though, this set mimics more than half of the vulnerabilities in our dataset. Due to the absence of any defined features to identify vulnerability-mimicking mutants, as our third contribution, this dissertation introduces Mystique, a deep learning approach that automatically extracts features to identify vulnerability-mimicking mutants. Despite the scarcity, Mystique predicts vulnerability-mimicking mutants with a high prediction performance, demonstrating that their features can be automatically learned by deep learning models to statically predict these without the need of investing any effort in defining features. Since our vulnerability-mimicking mutants cannot mimic all the vulnerabilities, we perceive that these mutants are not a complete representation of all the vulnerabilities and there exists a need for actual vulnerability prediction approaches. Although there exist many such approaches in the literature, their performance is limited due to a few factors. Firstly, vulnerabilities are fewer in comparison to software bugs, limiting the information one can learn from, which affects the prediction performance. Secondly, the existing approaches learn on both, vulnerable, and supposedly non-vulnerable components. This introduces an unavoidable noise in training data, i.e., components with no reported vulnerability are considered non-vulnerable during training, and hence, results in existing approaches performing poorly. We employed deep learning to automatically capture features related to vulnerabilities and explored if we can avoid learning on supposedly non-vulnerable components. Hence, as our final contribution, this dissertation proposes TROVON, a deep learning approach that learns only on components known to be vulnerable, thereby making no assumptions and bypassing the key problem faced by previous techniques. Our comparison of TROVON with existing techniques on security-critical open-source systems with historical vulnerabilities reported in the National Vulnerability Database (NVD) demonstrates that its prediction capability significantly outperforms the existing techniques. [less ▲]

Syntactic Vs. Semantic similarity of Artificial and Real Faults in Mutation Testing Studies

E-print/Working paper (2023)

Fault seeding is typically used in empirical studies to evaluate and compare test techniques. Central to these techniques lies the hypothesis that artificially seeded faults involve some form of realistic ... [more ▼]

Fault seeding is typically used in empirical studies to evaluate and compare test techniques. Central to these techniques lies the hypothesis that artificially seeded faults involve some form of realistic properties and thus provide realistic experimental results. In an attempt to strengthen realism, a recent line of re- search uses machine learning techniques, such as deep learning and Natural Language Processing, to seed faults that look like (syntactically) real ones, implying that fault realism is related to syntactic similarity. This raises the question of whether seeding syntactically similar faults indeed results in semantically similar faults and, more generally, whether syntactically dissimilar faults are far away (semantically) from the real ones. We answer this question by employing 4 state-of-the-art fault-seeding techniques (PiTest - a popular mutation testing tool, IBIR - a tool with manually crafted fault patterns, DeepMutation - a learning-based fault seeded framework and μBERT - a mutation testing tool based on the pre-trained language model CodeBERT) that operate in a fundamentally different way, and demonstrate that syntactic similarity does not reflect semantic similarity. We also show that 65.11%, 76.44%, 61.39% and 9.76% of the real faults of Defects4J V2 are semantically resembled by PiTest, IBIR, μBERT and Deep- Mutation faults, respectively. [less ▲]

Cerebro: Static Subsuming Mutant Selection

in IEEE Transactions on Software Engineering (2022)