Transfer Learning Strategies for Detecting Passive and GAN-Generated Image Forgeries with Pretrained Neural Networks

Image forgeries poses significant challenge in the area of digital forensics and security. Detecting forgeries created with more advanced image manipulation techniques especially those generated by Generative Adversarial Networks (GANs) has become critical task. Deep learning powered techniques appears to be most relevant in detecting digital forgeries. Hence, this study is an attempt to investigate the effectiveness of transfer learning using pretrained neural networks such as Alexnet, VGG16, Resnet50, InceptionV3, MobilenetV2 and EfficientNetB4 for detection of both passive and GAN-generated image forgeries. Experiments have been conducted on standard datasets like CASIA2.0 for passive forgeries and 140k real and fake faces for GAN-generated forgeries. Evaluated the efficacy of each pretrained model by considering the impact of fine-tuning and early stopping on overall detection performance. Performance analysis is done with the help of accuracy and loss curves as well as by considering precision, recall, f1 score and accuracy score values. Compared the model's ability in detecting both passive and GAN-generated image forgeries.