Deepfake Detection Using Meso4Net and Capsule Networks Through Facial Feature and Pattern Analysis
Nagalakshmi Pasupuleti, D. Venkata Lakshmi · IEEE Access · 2025
AI may now be expanded, and its technical potential is increasing every day. The quick expansion is causing risky issues. Complete alteration is taking place in the phony images and films. Images are being clicked and deep faked without anyone’s consent, which raises privacy concerns. This is highly targeted towards politicians and actors. The public also believes in deepfakes, and in these situations, individuals are unable to distinguish between genuine and fake. The purpose of this research is to determine which is right and which is not. The Facial Feature Analysis and Miniature Pattern Dissimilarity Verification model (FFA-MPDV), which combines meso4 for lightweight forgery detection with a capsule network to improve special feature retention, is part of the suggested model in this study. This makes it easier to spot modified characteristics, even subtle ones like lips, eyes, and facial expressions. Unlike traditional deepfake detection methods, which often struggle with subtle image modifications, the proposed FFA-MPDV model integrates meso4 and capsule networks, offering enhanced feature retention and sensitivity to fine-grained alterations. Additionally, the use of a spatial attention mechanism and FPN for multi-scale feature extraction ensures higher accuracy and faster identification. This unique combination significantly improves detection performance, achieving an impressive 97.3% accuracy, setting it apart from current state-of-the-art techniques and making it possible to identify which photographs are real and which are fraudulent in a matter of seconds.