Quantifying Deep Learning based Image Forgery Detection System: A Survey

The last few years have witnessed a high-pace rise in digital image forgery cases, where the forgers intend to manipulate images or videos to change information or even harass individual(s) for different reasons. On the other hand, the rising internet technologies and enabled applications like social media, e-commerce, and business communication which often use images or videos to circulate information have broadened the horizon for forgers to manipulate images. There are various types of image forgery, such as copy-move forgery, image-splicing, and resampling etc., which involve diverse patterns of manipulation and cause pixel-level disturbances that are challenging to detect, especially with the naked eye. Moreover, the advanced manipulation tools make it challenging to detect and classify image forgery cases. Localization is even more challenging task. Unlike traditional block-based or key-points based image forgery detection methods, which often makes use of the hand-crafted features to detect and localize forgery, the deep learning-based approaches can perform superior. The ability to learn local as well as contextual information from the input images make deep learning models more effective towards image forgery detection and classification. The improved deep models with their feature discrimination power strengthens to yield more accurate and reliable forgery detection and localization. This paper covers various systems for detecting and classifying image forgeries using deep learning. The different deep learning techniques such as the convolutional neural networks (CNN), recurrent neural networks (RNNs), auto-encoders, and other enhanced transfer learning methods such as AlexNet, MobileNet, ResNet, etc., designed towards copy-move forgery detection, splicing detection, and resampling detection etc. are discussed, with their strengths as well as limitations. Additionally, the different datasets available and their specifications are also discussed in details. The different challenges and allied scopes for improvement are also given in this manuscript. The concluding inferences along with the future scopes are also discussed in detail which can help researchers innovate and contribute superior towards image security and attack-resilience tasks.