Quiet Triggers, Loud Footprints: A Tri-Space Measurement Study of Backdoor Stealthiness.

arXiv:2610.19947v1 Announce Type: cross Abstract: Backdoor attacks on deep neural networks are commonly evaluated by attack success rate and benign accuracy, yet their stealthiness remains inconsistently defined and measured. Existing work often studies only o