Group-safe
All derivatives of one base document stayed in one split.
QUESTION
Threat model
The study classifies bona fide identity-document captures, printed reproductions, and documents recaptured from a screen. It does not evaluate identity matching, liveness guarantees, or fraud prevention in production.
PROTOCOL
Dataset decision
A proposed DLC-2021 + SIDTD four-class benchmark was rejected as the primary protocol because dataset source could deterministically reveal the composite class. Protocol A-Reduced instead uses 300 DLC-2021 samples balanced across three genuine classes.
Observed outcomes
Texture produced the highest observed test Macro-F1, driven primarily by PRINT performance. Edge improved over the semantic baseline overall, while SCREEN F1 fell in that configuration.
Evaluation discipline
Validation-only selection
Early stopping and checkpoint selection used validation Macro-F1.
Frozen test
The 44-sample test set was evaluated once after selection.
Interpretation
These are descriptive results from a reduced protocol and one primary seed. M4 and M5 also use projections and a deeper classifier than M3, so comparisons are not perfectly head-matched causal ablations.