Code Difference Guided Adversarial Example Generation for Deep Code Models

TL;DR

This paper introduces CODA, a novel adversarial example generation method for deep code models that leverages code differences to improve testing effectiveness and model robustness.

Contribution

It proposes a code difference guided approach (CODA) that reduces the search space and enhances adversarial example quality for deep code models.

Findings

CODA reveals 88.05% more faults than state-of-the-art techniques.

CODA improves testing efficiency and effectiveness.

Adversarial fine-tuning with CODA enhances model robustness.

Abstract

Adversarial examples are important to test and enhance the robustness of deep code models. As source code is discrete and has to strictly stick to complex grammar and semantics constraints, the adversarial example generation techniques in other domains are hardly applicable. Moreover, the adversarial example generation techniques specific to deep code models still suffer from unsatisfactory effectiveness due to the enormous ingredient search space. In this work, we propose a novel adversarial example generation technique (i.e., CODA) for testing deep code models. Its key idea is to use code differences between the target input (i.e., a given code snippet as the model input) and reference inputs (i.e., the inputs that have small code differences but different prediction results with the target input) to guide the generation of adversarial examples. It considers both structure differences and identifier differences to preserve the original semantics. Hence, the ingredient search space can be largely reduced as the one constituted by the two kinds of code differences, and thus the testing process can be improved by designing and guiding corresponding equivalent structure transformations and identifier renaming transformations. Our experiments on 15 deep code models demonstrate the effectiveness and efficiency of CODA, the naturalness of its generated examples, and its capability of enhancing model robustness after adversarial fine-tuning. For example, CODA reveals 88.05% and 72.51% more faults in models than the state-of-the-art techniques (i.e., CARROT and ALERT) on average, respectively.