Reproduce “TH-Bench: Evaluating Evading Attacks via Humanizing AI Text on Machine-Generated Text Detectors” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jingyi Zheng, Junfeng Wang, Zhen Sun, Wenhan Dong, Yule Liu, Xinlei He
The Hong Kong University of Science and Technology (Guangzhou) introduced TH-Bench, the first comprehensive benchmark to evaluate evading attacks on Machine-Generated Text detectors across effectiveness, text quality, and computational overhead. The benchmark revealed an "impossibility triangle" trade-off, where optimizing one dimension often compromises the others.
Key insights
- An inherent "impossibility triangle" trade-off exists among attack effectiveness, text quality, and computational cost, making simultaneous optimization challenging.
- No single evading attack uniformly excels; their effectiveness varies significantly across different detectors, datasets, and LLMs.
- Attacks often exhibit an inverse relationship: those highly effective in evasion tend to significantly degrade text quality or incur high computational overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.