Evaluating Deepfake Images: An Empirical Evaluation of Select Methods with Data Engineering

• Published in: Journal of Multimedia Information System Vol. 11; no. 4; pp. 229 - 240
• Main Authors: Echefu, Louis, Zhao, Qingsong, Chakrabarty, Subhajit
• Format: Journal Article
• Language: English
• Published: 한국멀티미디어학회 31.12.2024
• Subjects: 컴퓨터학
• ISSN: 2383-7632; 2383-7632
• DOI: 10.33851/JMIS.2024.11.4.229

In the rapidly evolving digital landscape, visual content—especially images and videos—plays a crucial role in online communication. However, the rise of deepfake technology, which employs deep learning techniques to create realistic manipulated media, raises significant ethical concerns due to its potential for misuse. This study conducts an empirical study of methods/tools for deepfake generation and detection, focusing on three prominent political figures: Vladimir Putin, Joseph Biden, and Narendra Modi. Using authentic images from the internet, we generated fake images using various deepfake tools and constructed a dataset comprising 600 real and 600 deepfake images. One of the key contributions of this paper was to integrate a data engineering approach. Among the models evaluated, the InceptionV3 model achieved the highest detection accuracy of 98.97%. Upon evaluating cross-datasets and combined datasets, we found that focused datasets improved model performance, emphasizing the importance of robust data engineering methodologies in addressing deepfake threats. This research contributes to the broader field of deepfake detection, with potential applications for other similar tasks. KCI Citation Count: 0