A Computer Vision Model to Identify Deepfakes to Combat Misinformation During Political Events

With rampant use of technology, deepfake videos have emerged as a critical issue in today's global landscape of rapid information dissemination, leading to new dimensions of influence and power. Political events frequently fall victim to misinformation; emphasizing that the authenticity of such propagated media is paramount. The aim of this paper is to utilize advanced AI and ML algorithms to combat the rising challenge of deepfakes - hyper-realistic videos that appear convincingly genuine to the human eye. This work utilizes Error Level Analysis to detect compression inconsistencies between frames and compares this approach with a modified CNN architecture. The modified InceptionResnetVl model that was created achieved a training accuracy of 78% and a validation accuracy of 75% when tested on the DeepFake Detection Challenge dataset. This was also tested with the Presidential Deepfake Dataset. A potential extension of this paper would be the integration of ELA with the modified CNN model significantly improving training efficiency and achieving a higher accuracy while also increasing computational simplicity.