MC-PDD Detects Pretraining Data in Black-Box LLMs via Masked Token Prediction

MC-PDD probes closed-source LLMs by masking high-specificity tokens and comparing prediction hit rates between candidate and reference corpora using statistical significance testing — no access to model logits required. This enables corpus-level membership inference against APIs like GPT-4 or Claude.