SERSEM: Membership Inference Attack Amplifies Memorization Signals in Code LLMs by Suppressing Boilerplate

arXiv·medium signal

Proposes SERSEM, a white-box membership inference framework for code language models that suppresses uninformative syntactical boilerplate to amplify specific memorization signals. Uses dual-signal methodology with character-level weighting. Critical for evaluating data contamination in code LLMs trained on potentially non-permissively licensed datasets — directly relevant as code model training data provenance becomes a legal and compliance concern.