Paraphrasing Attack Resilience of Various AI-Generated Text Detection Methods

This study evaluates the resilience of various AI-generated text detection methods against paraphrasing attacks, focusing on fine-tuned RoBERTa, Binoculars, and text feature analysis, as well as their ensembles with Random Forest classifiers. The findings reveal that ensembles including Binoculars achieve the highest detection performance but experience significant degradation under attack conditions. This highlights the trade-off between detection accuracy and resilience, raising critical concerns for practitioners regarding the reliability of current state-of-the-art detection techniques in combating AI-generated misinformation.