Performance Analysis of InceptionV3-RCNN and CGAN Models for Image Forgery Detection

Abstract

Due to the usage of advanced editing software and rising of AI-generated content, there seems to be the great problem with the forgery detection. This paper is a comparison of two deep learning models InceptionV3-RCNN and a Conditional GAN (CGAN) using CASIA v2.0 dataset which is benchmark for forgery detection. The result concluded that InceptionV3-RCNN network was highly effective on the augmented dataset with an accuracy of 98.61 while CGAN model had an accuracy of 76.52%. The training time for InceptionV3-RCNN is less and inferred time is more, whereas CGAN has taken more training time and less inferred time. On the whole, the two models bring out different strengths. The combined approach of InceptionV3-RCNN and CGAN be useful to improve the effectiveness of image forgery detection which could be explored in future work. As the results suggest, by integrating the strong classification properties of InceptionV3-RCNN and localization fine-grained properties of CGAN the combined features can be utilized.