MUSE-Net: Multi-stream Unified Subtle-artifact Evidence Network for Robust and Generalizable Deepfake Detection

Description

The dataset is designed to examine whether combining spatial, frequency-domain, and noise-related features can improve deepfake detection and manipulated-region localization. It contains real and manipulated facial images divided into training, validation, and test sets. Each sample is associated with an image path and a binary label indicating whether it is real or fake; pixel-level masks may be included for samples with localization annotations. The data are collected from established deepfake benchmarks and organized according to their original or defined experimental splits. Detection results are interpreted using metrics such as AUC, accuracy, precision, recall, and F1-score, while localization results are evaluated using IoU when valid pixel-level masks are available. Comparisons across datasets and ablation settings can be used to assess detection effectiveness, generalization ability, and the contribution of individual model components. The dataset itself does not establish a finding; observed differences in performance should be interpreted in relation to the dataset source, manipulation type, image quality, class distribution, and annotation completeness.

Institutions

• Liaoning Normal University

  Liaoning, Dalian

Categories

Licence