Xiangyue Liu

Title: Innovations of Xiangyue Liu in Backdoor Attack Methods

Introduction

Xiangyue Liu is a notable inventor based in Jiangsu, China. He has made significant contributions to the field of computer science, particularly in the area of security for classification tasks in code models. His innovative approach addresses critical challenges in the realm of machine learning and data integrity.

Latest Patents

Xiangyue Liu holds a patent for a "Backdoor attack method and system for classification task in code model." This patent outlines a method that includes several steps: first, collecting and preprocessing clean samples to obtain important variable names. Next, the variable names are classified according to label categories to create multiple trigger sets, followed by selecting target labels from the clean samples. The method then involves performing score calculations on the variable names in the trigger sets corresponding to the target labels. One important variable name is replaced with the variable name having the maximum C score in the clean samples to create poisoned samples. This process is repeated until the labels are changed to the target labels. Finally, triggers are randomly inserted into the poisoned samples to form negative samples, which are then used to perform an attack using an attack model based on the negative, poisoned, and clean samples.

Career Highlights

Xiangyue Liu is affiliated with Yangzhou University, where he continues to engage in research and development in his field. His work has garnered attention for its innovative approach to security in machine learning applications.

Collaborations

Xiangyue Liu collaborates with notable colleagues, including Xiaobing Sun and Yiran Xiao. Their combined expertise contributes to advancing research in their respective areas.

Conclusion

Xiangyue Liu's contributions to the field of backdoor attack methods highlight the importance of security in machine learning. His innovative patent reflects a deep understanding of the challenges faced in classification tasks, making significant strides in the field.