Loss-based
Model has lower loss on training data. Compute loss on candidate → below threshold → member. Simple but effective on undertrained models.
Advertisement
Shadow models
Train shadow models with known membership. Learn classifier on loss patterns. Apply to target model. Shokri et al 2017.
Advertisement
LLM-specific
Test perplexity on candidate. Compare to same-length random text. Lower relative → likely in training.
Real-world use
Determine if medical records, private conversations, copyrighted books were in training. Regulatory + legal implications.