Abstract:
"The rapid advances in Artificial Intelligence (Al) have led to the rise of deepfake technology, which presents significant issues for the integrity and authenticity of digital material. The ease with which fake images can be created has heightened the risks of misinformation, identity theft and other malicious activities. Addressing these issues necessitates the development of a practical deepfake detection approach to preserve the reliability of digital media and prevent potential misuse.
The author introduces a novel deepfake detection approach in this project that integrates advanced Deep Learning (DL) techniques. This novel approach capitalizes on the strengths of transfer learning by adapting a pre-trained MobileNetV2 model, enhancing its capabilities with custom layers specifically designed to identify and analyze images introduced by Generative Adversarial Networks (GANs). By integrating these custom layers, the image data can be analyzed more thoroughly and nuancedly, which helps the model identify subtle inconsistencies and deviations frequently found in images created through GANs. This architecture is complemented by an adaptive learning component that dynamically adjusts to the complexity and variability of the data, ensuring high sensitivity and specificity across a wide range of image qualities and resolutions.
This innovative approach achieved an impressive accuracy of 0.84 on the testing dataset, with a minimal loss of 0.3512, highlighting its effectiveness in discerning between real and fake images. The combination of high accuracy and advanced features highlights the model's robustness and adaptability, making it a significant contribution to deepfake detection."
