Forged Face Detection using ELA and Deep Learning Techniques

The paper presents real and fake facial image recognition using Deep learning and CNN. Image forgery recognition becomes a difficult task to find the authenticity of an image with the naked eye. This research aims to evaluate the working of different deep learning techniques in the novel “Real and Fake Face detection” dataset by Computational Intelligence Photography Lab, Yonsei University. For the detection of forged faces, the first step of the proposed method is image normalization for real and fake image recognition. Normalized images are then preprocessed using Error Level Analysis (ELA) and train to different pre-trained deep learning models. We finetune these models for categorization of 2 classes that are forged and real to evaluate these models' performance. From all tested models, VGG models give the best training accuracy of 91.97% and 92.09% on VGG-16 and VGG-19, whereas VGG-16 shows the good test set Accuracy using a smaller number of epochs, which is competitively better than all other techniques. Results of these models were evaluated using confusion matrix evaluation measures and compared with state of the art techniques.