WCBnet: Weighted convolutional block modelling of signed-value error levels for image-wise copy-move and splicing detection

Image manipulation which can easily generate hard-to-perceive fake information by image editing tools has become a threat of spreading visual mis/disinformation. With the speed and growth of such visual information presence in social media with respect to the current geopolitical affairs, tools for highly accurate verification of the authenticity of images are vital for AI-based fact checking. This work presents an efficient convolutional neural network (CNN) based approach for image manipulation detection. Our method, called WCBnet, starts with extracting learned features from the signed-value error levels (SEL) of compressed images on hierarchical convolution blocks. This is followed by adaptively concatenating, weighting and fusing these multi-level features by considering self-attention over all blocks according to different error levels corresponding to different manipulation types. We evaluate the performance of the proposed approach with respect to common manipulation datasets and compare with the state-of-the-art. WCBnet trained using around 2500 images of CASIA 2.0 dataset, resulted in the best F1-score for CASIA 1.0, Defacto, Coverage and Columbia datasets after fine-tuning by a small portion of those datasets. On average WCBnet improves the F1 score with respect to the second-best performing methods by 27.5%, 34.3%, 16.2% and 6.1% for these four datasets, respectively.