源代码处理任务中的深度学习模型对抗攻防研究综述OA

With the development of intelligent software,the application of deep learning models in source code pro-cessing tasks,such as defect detection and localization,has become increasingly widespread.But their lack of ro-bustness has also become increasingly evident.Many researchers have conducted in-depth studies on adversarial at-tack and defense methods for source code.However,existing surveys rarely summarize model characteristics from the perspective of source code task-specific properties,and there is a lack of systematic review and analysis of typic-al adversarial attack and defense methods such as model stealing,backdoor defense,and defensive distillation.Firstly,from the perspective of model architecture,we systematically outline deep learning models for source code processing tasks,and analyze their performance and adaptability under adversarial attack environments.Sub-sequently,we conduct a comprehensive review and classification of adversarial attack and defense methods for source code,and summarize the relevant benchmark datasets.Finally,we analyze the limitations of existing re-search and propose potential directions for future research.