An Image Forgery Detection Network with Edge and Noise Feature Fusion

Many current forged images are deceptively similar to the real ones, but deep learning techniques can still recognize subtle artifacts in them. However, many deep learning models are hard to understand, and their detection results may not be trustworthy. In this paper, we provide an image forgery detection network framework with interpretability, and design a lightweight image forgery detection network with edge and noise fusion named ENFIDNet. By slightly modifying the original network, the active output of what features and key regions the image forgery detection network model has focused on is ensured without compromising performance. The experimental results present that the network framework designed by us is not only able to actively provide a visual interpretation, but also offers a reference for the network modification.