Attention-based Fusion Network for Image Forgery Localization

With the trustworthiness of multimedia data has been challenged by editing tools, image forgery localization aims to identify regions in images that have been modified. Although the existing techniques provide reasonably good results for image forgery localization, with emerging new editing techniques, such models must be retrained and it is highly dependent on the real tampering localization maps. In this paper, we propose an attention-based fusion network that combines the RGB image and noise residual yielding excellent results. Noise residual is commonly regarded as camera model fingerprint, and forgery localization can be detected as deviations from the expected regular pattern. The model consists of three parts: feature extraction, attentional feature fusion, and feature output. The feature extraction module is used to extract RGB image features and noise residuals separately, and the attentional feature fusion module is used to suppress the high frequency components, supplement and enhance model-related artifacts by combining the aforementioned features. Finally, the last module generates images with one channel as the camera model fingerprint. In order to avoid dependence on tampering localization maps, the model is trained with pairs of image patches coming from the same or different camera sensors by means of Siamese network. Experiment results obtained from several datasets show that the proposed technique successfully identifies modified regions, improves the quality of camera model fingerprints, and achieves significantly better performance when compared to the existing techniques.