COMPARATIVE EVALUATION AND ANALYSIS OF DIFFERENT DEEPFAKE DETECTORS

Abstract

Deepfakes, synthetic media generated using deep learning, pose significant
risks to the integrity, security, and trust of information. Reliable detection
is therefore critical, yet existing models often fail when exposed to realworld
distortions such as compression, occlusion, and lighting variations. This
paper presents a comparative evaluation of deepfake detection models, including
XceptionNet, EfficientNet, MesoNet, and Vision Transformers, across multiple
benchmark datasets such as FaceForensics++, DFDC, Celeb-DF, and Wild-
Deepfake. Models are assessed not only under pristine conditions but also under
controlled distortions that reflect realistic deployment environments. The
results show that XceptionNet and the fine-tuned Vision Transformers achieve
the strongest accuracy and robustness, maintaining competitive performance
across domains, while MesoNet demonstrates computational efficiency but suffers
from reduced reliability under challenging conditions. EfficientNet provides
a balance between parameter efficiency and detection quality, but lags behind
in cross-dataset generalization. The findings highlight clear trade-offs between
robustness, efficiency, and deployment feasibility, emphasizing that lightweight
models are best suited for edge scenarios, whereas more complex architectures
remain preferable in cloud or high-resource environments. The study concludes
with open challenges and future research directions, including the integration of
multimodal cues, domain adaptation, and explainable detection frameworks, to
improve resilience against increasingly sophisticated deepfake generation techniques.