Lightweight Hybrid CNN-Transformer Architecture for Real-Time Image-Based Deep fake Detection on Resource-Constrained Systems

Contemporary face-manipulation detection draws on convolutional networks, vision transformers and hybrid designs, however published performance results are not directly comparable when evaluation datasets, hardware and inference protocols differ. This targeted narrative review analyzes 15 sources from the approved thesis synopsis: foundational image architectures, deepfake detection studies, two video benchmarks and deployment technologies. It distinguishes frame-level detection from video-level aggregation and temporal modelling, and considers whether published results establish feasibility on consumer-grade Intel hardware. The evidence supports the motivation for efficient local and global feature extraction, but does not justify a general claim that all accurate detectors require high-end GPUs or that existing hybrid systems are predominantly temporal. The review identifies a more specific evaluation need: matched accuracy and complete pipeline-latency measurements before and after model optimization. Therefore, a MobileNetV3 backbone with a shallow Transformer encoder and OpenVINO/NNCF deployment is therefore proposed as a research direction, not presented as a validated result.