Protocol choice strongly affects cross-family performance in feature-based hardware-Trojan detection

01 ABSTRACT

Using the Trust-Hub corpus and fixing the parser, 36 gate features, class weighting, model settings, threshold, and family aggregation, the study varies only the test split: pooled gates, one complete netlist held out, or every variant of one host family held out. Random forest F1/AP falls from 0.914/0.978 with pooled gates to 0.636/0.851 with one netlist held out and 0.460/0.577 with a host family held out; XGBoost falls from 0.946/0.976 to 0.464/0.544. The authors treat this as narrow evidence that sibling benchmark variants of one host circuit can inflate apparent transfer, and recommend family-aware holdouts plus all five family results alongside pooled scores; five families are too few for a population claim.

02 KEY FINDINGS

1. The test boundary strongly affects metrics: random forest F1/AP drops from 0.914/0.978 with pooled gates to 0.460/0.577 with a host family held out.
2. Direction is consistent across models: XGBoost falls from 0.946/0.976 to 0.464/0.544, while logistic regression loses AP but its fixed-threshold F1 is not monotonic.
3. Feature removal, repeated seeds, score normalization, and smaller samples change the gap size without reversing its direction, and aggregation choices also affect interpretation.