[4Xin2-98] The analysis of pretraining data detection on LLMs between English and Japanese
Keywords:Large Language Models, Membership Inference Attacks, Privacy Risk
The large amount of pre-training data used to build large language models (LLMs) may contain inappropriate data for training, such as copyrighted text or personal information. To solve this problem, the method for detecting the contents of LLMs pre-training data was proposed.
The existing method uses low token probabilities of a sequence for a determination.This method has been evaluated on LLMs trained English, and its effectiveness in LLMs trained Japanese has not been investigated.
In this study, we evaluated the effectiveness of the existing detection method on Japanese LLMs and compared it with the effectiveness on English LLMs. To this end, we constructed JAWikiMIA, a benchmark for detecting Japanese pre-training data.
We report that English LLMs achieve high AUC scores when the method uses the 20% of tokens from a sequence with the low token probability, while Japanese LLMs achieve high AUC scores when the method uses all tokens in a sequence.
The existing method uses low token probabilities of a sequence for a determination.This method has been evaluated on LLMs trained English, and its effectiveness in LLMs trained Japanese has not been investigated.
In this study, we evaluated the effectiveness of the existing detection method on Japanese LLMs and compared it with the effectiveness on English LLMs. To this end, we constructed JAWikiMIA, a benchmark for detecting Japanese pre-training data.
We report that English LLMs achieve high AUC scores when the method uses the 20% of tokens from a sequence with the low token probability, while Japanese LLMs achieve high AUC scores when the method uses all tokens in a sequence.
Please log in with your participant account.
» Participant Log In