Comparative Study on the Performance of Auto Encoders in Detecting Deep Fake Images

Due to the fast progress in deepfake technology, the world is experiencing fundamental shifts in how digital content is produced and consumed. Deepfakes, which rely on Artificial Intelligence (AI) for the creation of realistic but fake digital content, have considerable applications in different fields like entertainment, security, and social media platforms such as Facebook and Instagram. But it can be used in negative way also, recently it is used more for abuse and other illegal activities, so it is very important to differentiate which is fake and real. In this research, five different types of autoencoders namely, Convolutional autoencoders, Variational autoencoders, Hybrid Convolutional autoencoders, Contractive autoencoders and Vanilla autoencoder are compared on two different data sets namely “140k real and fake faces”, and “deepfake and real images”. These all models are part of the autoencoder family and follow same basic structure by using encoder and decoder. Encoder compress the input data into lower dimensional latent representation and decoder reconstruct the original data from the latent representation. Among the autoencoders which is implemented on the dataset “140k real and fake faces”, convolutional autoencoder achieve impressive accuracy of 98%, hybrid convolutional autoencoder achieves 97% and contractive autoencoder achieves 96% accuracy on classifying deepfake images from real images. In the dataset “deepfake and real images” the convolutional autoencoder achieve impressive accuracy of 90%, hybrid convolutional autoencoder achieves 93 %, and contractive autoencoder achieves 93 % accuracy on classifying deepfake images from real images.