Evolutionary Multi-objective Optimization for Contextual Adversarial Example Generation

Abstract

The emergence of the 'code naturalness' concept, which suggests that software code shares statistical properties with natural language, paves the way for deep neural networks (DNNs) in software engineering (SE). However, DNNs can be vulnerable to certain human imperceptible variations in the input, known as adversarial examples (AEs), which could lead to adverse model performance. Numerous attack strategies have been proposed to generate AEs in the context of computer vision and natural language processing, but the same is less true for source code of programming languages in SE. One of the challenges is derived from various constraints including syntactic, semantics and minimal modification ratio. These constraints, however, are subjective and can be conflicting with the purpose of fooling DNNs. This paper develops a multi-objective adversarial attack method (dubbed MOAA), a tailored NSGA-II, a powerful evolutionary multi-objective (EMO) algorithm, integrated with CodeT5 to generate high-quality AEs based on contextual information of the original code snippet. Experiments on 5 source code tasks with 10 datasets of 6 different programming languages show that our approach can generate a diverse set of high-quality AEs with promising transferability. In addition, using our AEs, for the first time, we provide insights into the internal behavior of pre-trained models.