A Proposed Multi-model Spatial, Temporal, and Frequency feature fusion Framework for Deepfake Video Detection

Abstract— The deepfake technology poses significant challenges to media authenticity, security, and public trust due to its rapid advancement. This paper reviews 40 recent research studies to identify gaps in existing deep-fake video detection methods. The current methods mainly rely on single-domain features and thus lack robustness across diverse datasets and real-world conditions. To address these limitations, a multi-model feature fusion architecture is proposed that integrates spatial, temporal, and frequency domain features to enhance detection accuracy and resilience. The research aims to design and evaluate this multi-domain fusion framework using systematic literature review and experimental implementation with advanced learning architectures. The key contributions include the development of a trained hybrid detection model demonstrating improved performance in varied scenarios, as well as providing insights into effective evaluation practices for building robust, generalizable deep-fake detection systems. This work advances the field by addressing critical gaps in generalization and multimodal feature integration. Keywords—Deepfake video detection, Hybrid machine learning, Multi-model feature fusion, Efficiency, Robustness