Preserving Spatial Information to Enhance Performance of Image Forgery Classification

As there are a huge range of powerful tools to edit images now, the need for verifying the authentication of images is more urgent than ever. While forgery methods are increasingly more subtle that even human vision seems hard to recognize these manipulations, conventional algorithms, which try to detect tampering traces, often pre-define assumptions that limit the scope of problem. Therefore, such methods are unable to encounter forgery methods in general applications. In this paper, we propose a framework that utilizes Deep Learning techniques to detect tampered images. Concretely, the MobileNetV2 network in [21] is modified so that it can be consistent to the task of image forgery detection. We argue that by remaining spatial dimension of early layers, the model is likely to learn rich features in these layers, and then following layers are to abstract these rich features for making a decision whether an image is forged. Besides, we also conduct a comprehensive experiment to prove those arguments. Experimental results show that the architecture-modified network achieves a remarkable accuracy of 95.15%, which surpasses others relying on the original architecture by a large margin up to 12.09%.