Author(s)
SABIHA SHABBIR CHIKTE, TAZMEEN RASHID AHMED KARBHARI, RADHA SUBODH ANERAO
- ISSN (P): 3139-8464
- Manuscript ID: 141158
- Volume: 2
- Issue: 10
- Pages: 740–749
Subject Area: Engineering, Science and Technology
Abstract
Face-swap deepfakes are now easy to make and hard to spot by eye, and they are already being used for misinformation, impersonation, and fraud. This paper reviews deep-learning methods for detecting such videos: benchmark datasets, frame-level CNN detectors, artifact-based and physiological-signal methods, and models that add temporal information. Detection is clearly possible, but results depend heavily on the dataset, the compression level, and the manipulation method, and most detectors return only a final label with no hint of where the manipulation is. Based on these gaps, we describe a prototype that classifies faces frame by frame, combines the frame scores into a Real/Deepfake decision with a confidence value, and shows the most suspicious frames on a web dashboard. We trained a baseline CNN, Xception, and EfficientNetB0 on the same data split and compared them. EfficientNetB0 gave the highest accuracy (59.61%) and ROC-AUC (0.6419) at about one-third of Xception’s inference time, so it was chosen for the prototype, although Xception had higher recall and F1-score. The scores are preliminary and leave clear room for improvement.