DEEPFAKE DETECTION USING VARIOUS DEEP LEARNING TECHNIQUES

Abstract

Deepfakes is a face swapping technique that allows anyone to change faces in a video with incredibly realistic results. But when used nefariously, this strategy can have a substantial influence on society, for example, by distributing bogus news or encouraging cyberbullying. As a result, the capacity to detect deepfakes is a critical concern. We address the subject of deepfakes detection in this research by detecting deepfakes in video frames. Existing research in the field of deepfake detection reveals that the increased obstacles given by new deepfake movies make detection approaches more difficult to detect. In this study we performed experiments using various SOTA architectures on DFDC dataset and then after comparing the performance of those both on accuracy and time to train ,found architecture which shows a perfect balance of accuracy and low computation time needed to train. The final proposed solution uses Efficient Net V2 as backbone in network to obtain competitive results.