一句话：提出自动分析结合人工复核的建模作业抄袭检测方法，对 ChatGPT 混淆抄袭的检测率与鲁棒性显著优于现有工作。
抄袭在计算机科学教育中普遍存在，大班课程难以人工逐一审查；而以 ChatGPT 为代表的大语言模型让抄袭者能轻易对答案进行混淆改写。此外，现有抄袭检测工具多面向代码，针对建模作业的方法很少，且普遍缺乏对混淆攻击的鲁棒性。本文提出一种面向建模作业的自动抄袭检测方法，将自动化分析与人工审查相结合，以兼顾检测能力与判定可信度。作者使用真实课程作业以及由 ChatGPT 混淆生成的抄袭样本进行评估。实验结果表明，与当前最先进方法相比，该方法对 AI 生成攻击的检测率显著更高，并且具备更广的鲁棒性。
Plagiarism is a widespread problem in computer science education, exacerbated by the impracticability of manual inspection in large courses. Even worse, tools based on large language models like ChatGPT have made it easier than ever to obfuscate plagiarized so-lutions. Additionally, most plagiarism detectors only apply to code, and only a few approaches exist for modeling assignments, which lack broad resilience to obfuscation attacks. This paper presents a novel approach for automated plagiarism detection in modeling assignments that combines automated analysis with human inspection. We evaluate our approach with real-world assignments and plagiarism obfuscated by ChatGPT. Our results show that we achieve a significantly higher detection rate for AI-generated attacks and a broader resilience than the state-of-the-art.
@inproceedings{SaglamHSB24a,
title = {Automated Detection of AI-Obfuscated Plagiarism in Modeling Assignments},
author = {Timur Sağlam and Sebastian Hahner and Larissa Schmid and Erik Burger},
booktitle = {ICSE},
year = {2024},
doi = {10.1145/3639474.3640084}
}