Benchmarking Deepfake Detectors: Comparative Analysis of Spatial, Temporal, and Hybrid Architectures

Abstract

Generative AI advancements are rapidly producing highly realistic deepfake videos, which raises major concerns for digital authenticity and security. This paper presents a comparative analysis of representative spatial, temporal, and hybrid deepfake detection architectures. Selected representative models include a lightweight MobileNetV2 + TinyViT hybrid approach; 2D Xception-based spatial model; 3D temporal video model; and the Knowledge-Guided Temporal Transformer (KGTT). Experimental results suggest that hybrid architectures provide the most balanced trade-off among the evaluated models, providing the best compromise between accuracy and computational efficiency. The KGTT model achieved frame-level accuracy exceeding 95%, while the MobileNetV2 + TinyViT hybrid achieved 80.46% accuracy with improved computational efficiency. Further, purely temporal models depend on large-scale, balanced datasets. Collectively, the findings show that integrating complementary feature domains is critical for developing robust and generalizable deepfake detection methods. This study provides a greater under-standing of the strengths and weaknesses of current approaches to deepfake detection, along with insights into promising areas for future research.