Lightweight Multi-Task Forgery Detection and Localization in Remote Sensing Imagery
Year: 2026
Type: preprint
Abstract: PRELIMINARY v1 preprint. This paper presents LMT-Forensics, a lightweight multi-task convolutional neural network (CNN) for forgery detection and pixel-level forgery localization in remote sensing imagery. The model combines an RGB branch with an SRM noise-residual branch and fuses them through a task-adaptive hyper-network, jointly predicting an image-level forgery label and a pixel-level forgery mask. This preliminary version reports a 2-epoch smoke test on EuroSAT (single seed): image-level AUC, F1 and accuracy of 1.0, pixel-level F1 of 0.9837 and pixel-level IoU of 0.9679. Full 5-seed results are pending, so the results reported here should be treated as indicative only.
Author: Altaf, Faheem
Institution: Islamia University of Bahawalpur
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225457868