RIFoL: A robust image forgery localization network for noisy images

The introduction of noise makes image forgery localization challenging. People can easily adopt additive Gaussian white noise and other types of noise to attack forged images. The poor robustness against noisy images has become a major problem in the practical application of image forgery localization techniques. To solve this problem, we propose an image forgery localization network, named RIFoL. RIFoL consists of a restoration module and a forgery localization module. The restoration module recovers the original forged traces from noise-attacked forged images. We design a forgery traces enhancement mechanism in the forgery localization module, including a Spatial Feature Enhancement Module (SFEM) and a Multi-Attention Feature Fusion Module (MAFM). The SFEM employs a multi-scale attention mechanism, incorporating channel and spatial attention information into the feature enhancement module. The MAFM aggregates features from two scales through spatial and channel pathways. Our method demonstrates superior performance in detecting and localizing noise-attacked forged images compared to existing state-of-the-art approaches.