Machine Learning Augmented Fuzzing

Abstract

The proposed thesis introduces cutting edge Machine Learning (ML) tools into Search Based Software Engineering (SBST). The contribution is three-fold. The first is an ML driven property targeting search strategy. It uses a deep neural network to process execution trace information to yield a likelihood score of a presence of a crash, which is in turn used as a fitness function for search. This method clearly outperforms the baseline search technique. The second contribution is a method for defining a property agnostic search landscape. This is achieved by training an autoencoder on a corpus of execution traces to produce a "latent space" representation. The expectation is to observe a tendency for arbitrary properties of executions to group in distinct regions of the latent space. Location in this space would in turn be used to direct an SBST process. The third contribution is to augment an automated tool with a generative model. The intention is to produce approximately valid input seeds that would target desired locations of a fitness landscape. These contributions will provide novel ideas for future research in the intersection of SBST and ML.