Multiple Manipulation Detection in Images Using Frequency Domain Features in 3D-CNN

In digital era, image can be easily forged by multiple manipulations using advance editing tools, such that truthfulness of that image cannot be identified by human eye. Many approaches have been proposed for the detection of these forged images. However, the performance of these approaches is quite better for large resolution and uncompressed images, whereas they fail for small-sized highly compressed images. To address this issue, a novel DCT-3DCNN architecture is proposed for multiple manipulation detection. The proposed DCT-3DCNN is constructed by stacking the DCTs of four residuals (Average filtering residuals, Gaussian filtering residuals, Laplacian filtering residuals and median filtering residuals) along depth-wise. The four DCTs are more capable to extract the manipulations traces in an image. These traces are fed into 3D-CNN to learn the low to high level features of multiple manipulations. Thus, the features are combined to classify the forged and pristine images. The performance of the proposed DCT-3DCNN is supported by exhaustive experiments for binary classification and multi- class classifications. Experiments are conducted on five (UCID, RAISE, BOSSBase, BOWS2 and NRCS) databases. The robustness of the proposed network is also evaluated for the detection of bilateral filtering on images. For binary classification, the improvement ratio (%) between the proposed (DCT-3DCNN) and state-of-the-art methods (MFR-CNN, RF-CNN) is 4–5%, while for bilateral filtering the improvement ratio (%) is 8% in comparison with the state-of-the art method RF-CNN. The proposed network achieves 14% improvement in detection accuracy for multi-class classification as compared to the RF-CNN.