Generalization and Cross-Dataset Robustness in Deepfake Detection: An Enhanced XceptionNet Approach

Abstract: Maintaining the integrity of digital visual media in the age of generative AI requires robust, automated frameworks capable of identifying sophisticated facial manipulations. This paper presents the design and evaluation of a specialized benchmarking framework for Deepfake detection, developed using an XceptionNet architecture to analyze and improve cross-dataset generalization. The proposed system compares a baseline detector against an enhanced variant (V2) that integrates strategic data augmentation and domain-specific fine-tuning to bridge the performance gap between source and target datasets. To ensure evaluation rigor, the framework supports multi-seed experimentation with deterministic sampling, enabling statistically grounded comparisons across independent training runs. This reproducible design eliminates variance-driven conclusions and strengthens the reliability of cross-dataset generalization findings. . Researchers are provided with a diagnostic toolkit that utilizes ROC/AUC analysis and bootstrap statistics to ensure the reliability and significance of detection metrics. Experimental results indicate that the targeted training interventions significantly enhance the model's ability to maintain high accuracy across unseen distributions without degrading performance on original training data. This research demonstrates how a systematic benchmarking approach can diagnose model weaknesses and provide a reproducible pathway toward developing more resilient and "wild-ready" Deepfake detection systems.