The rapid advancement of generative artificial intelligence and deepfake technology poses a profound threat to digital security, social trust, and the integrity of judicial evidence, necessitating the development of robust, adaptive, and interpretable verification Model.
This thesis proposes an optimized Deep Learning model based on a customized Convolutional Neural Network (CNN) architecture designed specifically to overcome the inherent limitations of traditional “black-box” systems and mitigate the lack of generalization across diverse domains.
Evaluated on the Deepfake Detection Challenge (DFDC) dataset, the proposed model achieved a high performance internal accuracy of 98.0% and an Area Under the Curve (AUC) of 0.99.
To establish scientific validity, a rigorous statistical verification was conducted, yielding a 95% Confidence Interval of [97.36% – 98.64%] and a p-value < 0.001. To address the critical challenge of the “domain gap” and ensure generalization capability, a strategic transfer learning and fine-tuning methodology was implemented and cross-validated on the external Celeb-DF v2 dataset, which effectively enhanced cross-dataset accuracy from an initial 58.0% to a robust 89.0%.
A core novelty of this study lies in the integration of Explainable AI (XAI) techniques—specifically SHAP (SHapley Additive exPlanations) and Grad-CAM (Gradient-weighted Class Activation Mapping)—providing granular pixel-level and spatial visual explanations that localize facial manipulation regions, thereby establishing forensic transparency. Furthermore, the model demonstrates high computational efficiency, achieving an inference speed of 4.65 Frames Per Second (0.21 seconds per image) entirely on CPU resources, validating its readiness for real-time forensic screening in resource-constrained operational environments.
Ultimately, this research provides a highly balanced framework that bridges the gap between high-precision classification, statistical reliability, and cross-dataset robustness, offering a transformative tool for digital forensics and judicial investigations.
Keywords: Deepfake Detection, Convolutional Neural Networks (CNN), Transfer Learning, Explainable AI (XAI), SHAP, Grad-CAM, Forensic Integrity, Statistical Validation, Domain Gap.
