JAIT Three-Dimensional Convolutional Approaches for the Verification of Deepfake Videos: The Effect of Image Depth Size on Authentication Performance

ICIET 2023 -T217

Abstract

Deep learning has proven to be particularly effective
in tasks such as data analysis, computer vision, and human
control. However, as this method has become more advanced,
it has also led to the creation of DeepFake video sequences and
images in which alterations can be made without immediately
appealing to the viewer. These technological advancements have
introduced new security threats, including in the field of education.
For example, in online exams and tests conducted through
video conferencing, individuals may use Deepfake technology to
impersonate another person, potentially allowing them to cheat
by having someone else take the exam in their place. To address
these issues, several detection approaches have been proposed,
including systems that use both spatial and temporal features.
However, existing approaches have limitations in terms of detection
accuracy and overall effectiveness. This paper offers our
technique for combining convolution neural network (CNN) with
temporal analysis for effective Deepfake detection. We compared
and analyzed different 3-D CNN-based model approaches and
varied sequence lengths of facial photos used in conjunction with
the 3-D CNN models in the FaceForensic++ datasets. Our findings
demonstrated that the model of 3-D CNN using 16 sequential face
images as input provided up to 97.3 % accuracy in detecting
Deepfakes.