A System For Identifying Synthetic Images Using Lstm: A Deep Learning Approach

摘要

In the current scenario Generative Adversarial Network (GAN) is generating more exhilaration in various fields with an amazing growth of it can be seen over a few years. It is very much successful in generating synthetic images over natural images. These are unsupervised neural networks that are capable of creating new image samples based on the training process they have adapted from the information that has been fed to them. On the other hand, Long Short Term Memory (LSTM) is one type of Recurrent Neural Network (RNN) mainly used in the domain facing sequence prediction issues. In this paper, the GAN is considered a Generator, and the LSTM is considered a Discriminator. The work of the generator is to produce synthetic images out of random samples. Based on the finetune training, it can produce a perfect fake image that is difficult to identify as a real one. The same is fed to the LSTM network along with the real images, and the finetune training is performed to get more perfect synthetic images. Both facial datasets, as well as abstract art dataset available opensource, is taken for training and testing. From this research, it is proven that Generative Adversarial Network (GAN) and Long ShortTerm Memory (LSTM) are the networks utilized, and the accuracies were found to be 58.53% and 72.68%, respectively, which explicitly proves that synthetic images are more clearly identified by the LSTM over GANs.