Blur Sensitivity and Generator Dependence of Fingerprint Alteration Detectors Trained on SOCOFing
M. K. S. Likhon, M. M. Sarker
Manuscript prepared for Image and Vision Computing (Elsevier), 2026
Ten model variants reach 0.996–0.998 macro-F1 on SOCOFing, with no gain from fusing minutiae. Yet a one-pixel Gaussian blur makes eight of nine image models call 93–97% of genuine prints altered, and the same alteration types from an independent generator are mostly missed (macro-F1 0.38–0.42). Blur-augmented training cuts the false alarms from 94% to 3% but does not close the generator gap.
Data Audit and Revised Baselines for Structural Damage-Type Classification on PEER Hub ImageNet
M. K. S. Likhon, M. M. Sarker
Manuscript in preparation, 2026
An audit of the damage-type task of PEER Hub ImageNet finds 91 near-duplicate pairs in the training set, 30 with conflicting labels, and 23 test images with a near-copy in the training set. Under duplicate-aware cross-validation with no tuning on the test set, an ensemble of current image models reaches 77.4% test accuracy, against 72.4% for the best published result on the official split.
Data auditNear-duplicatesLabel noisePEER Hub ImageNet
No publications match “”.
Research write-ups that are not papers yet are on the projects page.