Automatic Test Data Generation for a Given Set of Applications Using Recurrent Neural Networks

To address the problem of automatic software testing against vulnerabilities, our work focuses on creating a tool capable in assisting users to generate automatic test sets for multiple programs under test at the same time. Starting with an initial set of inputs in a corpus folder, the tool works by clustering the inputs depending on their application target type, then produces a generative model for each of these clusters. The architecture of the models is falling in the recurrent neural network architecture class, and for training and inferencing the models we used the Tensorflow framework. Online-learning is supported by the tool, thus models can get better as long as new inputs for each application cluster are added to the corpus folder. Users can interact with the tool similar to the interface used in expert systems: customize various parameters exposed per cluster, or override various function hooks for learning and inferencing the model, with the purpose of getting finer control over the tool’s backend. As the evaluation section shows, the tool can be useful for creating important sets of new inputs, with good code coverage quality and less resources consumed.