MC-PDD: Masked Corpus-Level Pretraining Data Detection for Black-Box Large Language Models
Researchers have developed a new method called MC-PDD to detect if specific datasets were used in the pretraining of large language models, even for black-box, closed-source models. This technique, inspired by masked language modeling, masks tokens and assesses the model's prediction accuracy to determine data inclusion. MC-PDD offers performance comparable to existing methods while operating solely through standard API access, enabling applications like model auditing and copyright verification. AI
IMPACT Enables auditing of LLM training data and verification of data copyright using only API access.