SynShield: Hybrid Detection of Synthetic Medi...

SynShield: Hybrid Detection of Synthetic Media with EfficientNet and MesoNet

Author Name : Richa Sharma, Tripti Sharma

DOI: https://doi.org/10.56025/IJARESM.2025.1307250033

ABSTRACT The proliferation of deepfakes—highly realistic manipulated media generated using deep learning—has raised serious concerns regarding misinformation, digital identity theft, and public trust. This paper proposes DeepFence, a hybrid deepfake detection framework that synergistically combines EfficientNet and MesoNet to capture both semantic and mesoscopic inconsistencies in facial videos. EfficientNet extracts high-level global features, while MesoNet tries to target local texture anomalies mostly introduced during forgery. The features are supplied to a binary classifier trained with the FaceForensics++, Celeb-DF v2, and DFDC datasets after being fused using global average pooling. The suggested model outperforms standalone systems and exhibits good generalisation to invisible alterations, with a detection accuracy of 94.7% on FF++. The model architecture is further validated by theoretical foundations based on generalisation bounds, multiple kernel learning, and feature space complementarity. DeepFence provides a scalable and effective way to combat the risks associated with generative media technologies.