Generalization of Deep Neural Networks in Image Manipulation Detection

In today's digital landscape, the widespread availability and simplicity of image manipulation software has contributed to a significant increase in the dissemination of manipulated images. Combined with the rapid spread of controversial content on social media platforms, there exists a critical need to protect individuals and institutions from malicious disinformation campaigns and unjust cancel culture. This calls for robust methods of detecting and combating image manipulation. This thesis focuses on analyzing the capabilities of two state-of-the-art deep convolutional neural networks, CAT-Net and MVSS-Net, in the context of image manipulation detection. By evaluating the applicability of these networks, the research aims to identify and measure the existing gap in safeguarding against the harmful effects of manipulated images. Building upon prior research, the thesis introduces three key experiments to assess the generalization ability of CAT-Net and MVSS-Net. The first experiment focuses on calculating the receptive field size of the networks, shedding light on their maximum capabilities in the face of improving image resolution and quality. The second experiment investigates the networks' sensitivity to manipulations, suggesting sensitivity to minute alterations as small as 100 pixels in total. These experiments provide valuable insights into the networks' strengths and limitations, contributing to a comprehensive understanding of their effectiveness in real-world scenarios. The final experiment employs a recently theorized concept called the generalization interval to evaluate the networks' performance in detecting complex, real-world image manipulations. The findings from this experiment reveal that neither network exhibits accurate detection capabilities under the forgiving conditions of easy, human-detectable manipulations. However, CAT-Net's architecture demonstrates promising potential for future solutions, displaying commendable localization performance while results highlight strong memorization characteristics in the model. In conclusion, this thesis addresses the pressing need for effective image manipulation detection methods in the era of widespread access to image manipulation software and the rapid dissemination of controversial content. By evaluating the generalization ability of CAT-Net and MVSS-Net, the research contributes to the ongoing efforts to protect individuals and institutions from the adverse and unjust effects of disinformation. The findings emphasize the importance of advancing the capabilities of deep neural networks and developing more robust solutions to combat image manipulation in various real-world scenarios.