Robust forgery detection and localization research based on dynamic reinforcement and enhanced sliding window

Aiming at the problems of poor robustness of forgery detection, difficulty in capturing global information, and suboptimal training with fixed parameters brought about by the post-processing operations such as compression and scale stretching of images on social platforms such as WeChat and Weibo, we propose a method for robust forgery detection and localization based on dynamic reinforcement and augmented sliding windows. First, we use adversarial training to generate images with noisy perturbations. Then, we propose an end-to-end detection and localization network with an enhanced sliding window of a U-Net-like structure, which can flexibly focus on the global information of each region of the image as well as the regions between the regions, and use the U-decoder to extract the correlation information of the image context at multiple scales. In addition, we propose a dynamic reinforcement algorithm for optimized training, which adaptively adjusts the hyper-parameters of the network during the network training process. The experimental results show that compared with the current mainstream models, the model in this paper holds an advantage in the test results on different OSN platforms, and for Facebook, which is rich in post-processing operations, this method improves the F1 and IoU by 4% and 1% on the DSO test set, 11% and 5% on the NIST16 test set, and 16% and 7% on the CASIAv1 test set, respectively.