Neural architecture search for detection of deepfakes

Abstract

DeepFakes has become a social problem by slandering the public image of people. The current and specialized DeepFake detection methods use Hand-Designed Deep Neural Networks as backbone, which requires a lot of designing effort. We propose to use a Neural Architecture Search algorithm for DeepFake detection, which aims to overcome the drawbacks of the Hand-Designed approaches, and has not been proved yet on DeepFake detection. Additionally, in this work is analyzed the robustness of the Architecture Search process. We yield comparable results to DeepFake detection state-of-the-art, that are highly specialized, by adapting Progressive Differential Architecture Search to DeepFake detection for the first time, using novel techniques such as data augmentation, multi-label classification and an architecture search process in the particular domain to improve in terms of performance. Finally, We have found that the data requirements to obtain a stable architecture are not very high.