Publication Date

Fall 12-22-2020

Degree Type

Master's Project

Degree Name

Master of Science (MS)

Department

Computer Science

First Advisor

Teng Moh

Second Advisor

Melody Moh

Third Advisor

Robert Chun

Abstract

Advances in generative models and manipulation techniques have given rise to digitally altered videos known as deepfakes. These videos are difficult to identify for both humans and machines. Typical detection methods exploit various imperfections in deepfake videos, such as inconsistent posing and visual artifacts. In this paper, we propose a pipeline with two distinct pathways for examining individual frames and video clips. The image pathway contains a novel architecture called Eff-YNet capable of both segmenting and detecting frames from deepfake videos. It consists of a U-Net with a classification branch and an EfficientNet B4 encoder. The video pathway implements a ResNet3D model that examines short clips of deepfake videos. To test our model, we run experiments against the Deepfake Detection Challenge dataset and show improvements over baseline classification models for both Eff-YNet and the combined pathway.

Available for download on Wednesday, December 22, 2021

Share

COinS