LoMime: Query-Efficient Membership Inference Attacks via Model Extraction in Label-Only Settings

Researchers present LoMime, a membership inference attack that determines whether specific data points were used to train a machine learning model while requiring only the model's predicted labels. The approach uses model extraction to improve query efficiency, relaxing common assumptions such as access to confidence scores, shadow models, or the training data. The work highlights privacy risks for deployed models under realistic, label-only access conditions.