Protocol effects on feature-based hardware-Trojan detection across Trust-Hub families

Trust-Hub is widely used to evaluate gate-level hardware-Trojan detectors, and many of its netlists are Trojan variants of the same host circuit. Common gate and netlist splits can therefore place nearly identical benign logic in both training and test data. We quantify this effect with a fixed 36-feature pipeline and three standard classifiers. The study covers 49,124 gates from 16 netlists grouped into five basecircuit families. We compare pooled cross-validation with leaveone-circuit-out (LOCO) and leave-one-base-family-out (LOFO) evaluation. Random-forest F1 is 0.914 under pooled validation, 0.636 under LOCO, and 0.460 under LOFO; average precision changes from 0.978 to 0.851 and 0.577. XGBoost shows a similar decline. The gap persists across feature ablations, repeated seeds, family-level results, and a reduced-sample check. These measurements show how strongly benchmark reuse can affect
the apparent generalization of a gate-level detector.