Paraphrasing Attacks on Text Detection Systems: A Survey

In this article, we explore the use of watermarking to combat plagiarism in machine-generated text, particularly in the context of news articles. We discuss two datasets used in our experiment: C4 and Wikitext. Our approach focuses on token-level paraphrasing attacks to evaluate the robustness of the models. We analyze the effectiveness of watermarking on textual quality and report P-SP, diversity, and coherence scores.
Watermarking is a crucial technique in detecting plagiarism in machine-generated text as it adds a unique identifier to the generated text, making it traceable back to its original source. However, recent advancements in machine learning have made it easier for attackers to manipulate the watermarked text, making it indistinguishable from human-authored text. To address this issue, we propose a novel approach that combines token-level paraphrasing attacks with other methods to enhance the robustness of watermarking models.
Our findings show that our proposed method significantly improves the detection rate of plagiarism in machine-generated text compared to existing approaches. We also observe an improvement in the textual quality of the generated texts, as measured by P-SP, diversity, and coherence scores. Furthermore, we provide samples of watermarked texts in the Appendix A for a more comprehensive understanding of our approach.
In conclusion, our study demonstrates the effectiveness of watermarking machine-generated text to combat plagiarism. By combining token-level paraphrasing attacks with other methods, we can significantly improve the robustness of watermarking models and enhance their ability to detect plagiarism in machine-generated text. Our approach has important implications for various applications, including news article generation and language translation.