Deceiving neural source code classifiers: Finding adversarial examples with grammatical evolution

This work presents an evolutionary approach for assessing the robustness of a system trained in the detection of software vulnerabilities. By applying a Grammatical Evolution genetic algorithm, and using the output of the system being assessed as the fitness function, we show how we can easily change the classification decision (i.e. vulnerable or not vulnerable) for a given instance by simply injecting evolved features that in no wise affect the functionality of the program. Additionally, by means of the same technique, that is by simply modifying the program instances, we show how we can significantly decrease the accuracy measure of the whole system on the dataset used for the test phase. Finally we remark that these methods can be easily customized for applications in different domains and also how the underlying ideas can be exploited for different purposes, such as the exploration of the behaviour of a generic neural system.