Application of spatial and Wavelet transforms for improved Deep Fake Detection

Abstract

Deepfake technology has been controversial for the past recent years. It has been considered a double-edged sword for what it has the capability of doing. As much as deep fake technology can be beneficial in many situations, it has caused a global concern over the safety of individuals. It can be used to create inappropriate content to blackmail someone, or it can be used to fabricate fake news and cause chaos in society. Therefore, developing algorithms to detect whether any content is authentic or tampered with is crucial to protect people. With the evolution of Generative Adversarial Networks (GANs), many deepfake tools have evolved to create non-distinguishable content. This new evolution has raised the flag of the necessity to get matters into control. In this research, various new methods to detect deep fakes are proposed using the frequency analysis (Fourier transform and wavelet transform) of the frames of the videos to discriminate the footage achieving an accuracy of 92.24% using a novel CNN-LSTM model trained on Wavelet transform data. This research argues that the wavelet transform can hold information and features that can outperform training on spatial images on a custom dataset that combines many various datasets.