Real or Fake: a Technique to Detect Manipulated Images

Numerous images arise everyday and have largely substituted texts to become the most significant sources of acquiring information. However, the rapid development of image editing tools make it difficult to judge whether an image authentic or which part of an image is forgery. From powerful softwares like Photoshop and face editing apps to advanced technologies like GAN, it is more and more difficult and time-consuming for human eyes to distinguish them. This paper use CNN to detect real and fake images. We focus on the photoshopped images since it is convenient and fast to edit images using such tools compared to more complex methods like generating adversarial images. These editing methods include splicing, copy-move, removal, enhancement, etc. and they are the most commonly utilized ones. Firstly we separate the two functions and use the class model for real/fake classification and the location model for fake part segmentation. An image will be firstly tested whether it is real or fake and if it is classified as fake, the location model can detect the specific fake regions. Later we merge these two models into one model to provide an end-to-end solution. Therefore, both the training and test can be completed more compactly. Finally, experiments are conducted on several different datasets and our models achieve state-of-the-art or comparable results on all of them, demonstrating that our model is capable of detecting and localizing forgery regions of images editied by various methods. Additionally, models are tested on images manipulated by advanced neural networks, further uncovering the generalizability and insufficiency of our models.