Increasing the Robustness of the Fine-tuned Multilingual Machine-Generated Text Detectors
Abstract
AbstractSince the proliferation of LLMs, there have been concerns about their misuse for harmful content creation and dissemination. Recent studies justify such fears, providing evidence of LLM vulnerabilities and high misuse potential. It has been shown that humans struggle to reliably distinguish high-quality machine-generated text from authentic human-written text. Therefore, it is crucial to develop automated means to accurately detect machine-generated content. While multiple detectors of machine-generated texts have been researched so far, a known issue is their low generalizability to unseen generators, domains or languages. This work addresses the problem by proposing a robustness-oriented LLM training methodology for the detection task, making the detectors more robust against obfuscation and more generalizable to out-of-distribution data. The fixed-size ablation study confirms that the gains are not simply a function of more data.