LIED: A Lightweight and Ensemble learning approach for fake face Detection

For many years, machine learning problems have primarily been driven by the availability and quality of data. Being the key, data has been equally vulnerable and got a savior in the form of generative adversarial networks (GANs) which opened the floodgates for generating almost any type of real, yet synthetic data. Human face became one of the initial victims to this superior technology, where in highly realistic and convincing fake content is generated using deep learning technologies or ‘‘DeepFakes’’. Taking a giant leap forward from manipulating facial attributes, to be now able to swap expressions seamlessly and even generate new (non-existent) synthetic faces poses a grave threat not only to chosen few, but for the entire society. This upshoot has been reciprocated with significant efforts and investments for its detection, but the techniques are often marred with either lower accuracies, or, higher computation costs. This is where convolutional reservoir networks (CoRN) come to rescue owing to their lightweight nature, able to do ensemble feature extraction and its generalization ability. This paper investigates, implements and demonstrates the application of CoRN based architectures to the task of human fake face detection. The steep performance improvements as evident from our results further ratify the effectiveness of this approach, which is also shown to perform exceedingly well against smaller datasets.