VERITAS: A Novel Framework for Forensic Deepfake Detection with Visually Interpretable XAI

Abstract

"The rise of deepfake technology presents a formidable challenge to various sectors, including media, politics, and cybersecurity, as it enables the creation of highly convincing synthetic media content that can deceive and manipulate audiences. To combat this threat, this research project focuses on developing an novel deepfake detection framework that leverages advanced technologies and interdisciplinary insights.

The research proposes a novel approach to deepfake detection that integrates visually interpretable eXplainable Artificial Intelligence (XAI) methods. The proposed framework encompasses state-of-the-art machine learning models, such as EfficientNetB0 pre-trained models, and combined XAI techniques to enhance the interpretability of detection results. Furthermore, the project explores the creation and curation of deepfake datasets, including the Faceforensics++ dataset and DFDC dataset, to facilitate model training and evaluation.

The initial results of the deepfake detection framework showcase its efficacy in discerning between manipulated and authentic media content with high accuracy. The model achieves a training accuracy of 90.73%, validation accuracy of 80.54%, and testing accuracy of 85.31%, demonstrating its robust performance across different stages of the training process. The framework exhibits a balanced trade-off between precision and recall, with a precision score of 0.877 and a recall score of 0.845. The F1-Score score of 0.821 reflects the model's strong discriminatory power and its capacity to effectively separate genuine from manipulated media content. Overall, the initial results validate the effectiveness of the proposed deepfake detection framework and highlight its potential to address the challenges posed by synthetic media manipulation. "