Manuscripts & reports

Joint Prediction of Ultra-High-Performance Concrete Properties Using a Feature-Tokenizer Transformer with Masked Multi-Task Learning

R. Tanvir*, M. K. S. Likhon*, et al.

Submitted to Applied AI Letters (Wiley), 2026 · Under review

A feature-tokenizer Transformer predicts several properties of ultra-high-performance concrete at once. Masked multi-task learning lets the model train on samples where only some properties are labelled.

Out-of-Focus Is Not Obliteration: A Perturbation Analysis of Fingerprint Alteration Detectors on Synthetic Benchmarks

M. K. S. Likhon*, M. M. Sarker*

Manuscript in preparation, 2026

Detectors reach 0.998 macro-F1 on the synthetic SOCOFing benchmark. A perturbation analysis shows they learn the traces of how the images were edited, not the alteration itself.

Does Geometry Help? A Leakage-Audited, Negative-Result Study of Cross-Attention Fusion and Taxonomy-Consistency Loss for Post-Earthquake Damage-Type Classification

M. K. S. Likhon*, M. M. Sarker*

Manuscript in preparation, 2026

A study of cross-attention geometry fusion and a taxonomy-consistency loss on PEER Hub ImageNet, audited for data leakage. Geometry fusion performs no better than a control that uses shuffled geometry.

Design of an Arrhythmia Classification Algorithm Using 2-D Convolutional Neural Networks

M. K. S. Likhon

B.Sc. thesis, Ahsanullah University of Science and Technology