Black-box auditing of quantum models: Lifted differential privacy based on quantum canaries

The researchers propose the first black-box privacy audit framework for quantum machine learning (QML), based on lifted quantum differential privacy, using quantum canaries (offset-encoded quantum states) to detect memorization during training and quantify privacy leakage. The framework establishes a mathematical connection between canary offset and trace-distance bounds, derives an empirical lower bound on privacy budget consumption, and bridges the gap between theoretical guarantees and practical verification. Evaluations on simulated and physical quantum hardware verify the framework's effectiveness in measuring the actual privacy loss of QML models.