Year: 2024
Type: dissertation
Abstract: The recent evolution of deepfake technology as a result of intelligence growth is a development that has come with challenges to its security systems such as those that used facial recognition. This research aims at identifying deepfakes using a CNN model which is created using MobileNetV2 architecture. The face and the model were trained using the videos which belong to FaceForensics++ dataset featuring with various manipulated videos that aimed at covering different kinds of deepfakes. The aim of this work was to improve the detection performance and stability as a means of improving the real time applications in the security systems. The framework encompassed data preprocessing steps including frame extraction resizing normalization and augmentation the training and optimization employing the Adam optimizer was also involved. It was assessed using some performance indicators such as accuracy, precision, recall and the F1 measure. The outcomes established higher results of the detection accuracy with detecting accuracy rate of 87 percent and performance in relation to the different tests object sample with different video resolution and ways of deepfakes creation. In doing so this research offers a practical solution as to how organisations can enhance the existing frameworks thus making them better prepared to handle risk related to Artificial Intelligence technologies. It does not only fulfill the needs of security systems but also opens the doors to the new development of detecting more advanced deepfakes. More work will be done on increasing the data set to make the model more responsive as well as improving real time functionality for use cases.
Author: Sharma, Himanshu
Institution: National College of Ireland
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7056357516
