Improved Hybrid Convolutional Neural Network Approach Based on Deep Learning for Deepfake Detection

Abstract

The rapid advancement of deep learning has enabled the creation of highly realistic deepfake media, raising concerns in cybersecurity, misinformation, and digital forensics. Traditional deepfake detection approaches primarily rely on convolutional neural networks (CNNs), which are often vulnerable to adversarial attacks and struggle to generalize across datasets. This study proposes a hybrid deepfake detection framework integrating ResNet50, MesoNet4, and an Eye Movement Analysis module to improve detection accuracy and robustness. ResNet50 is chosen for its ability to extract high-level spatial features, while MesoNet4 is optimized for detecting low-resolution manipulations. Additionally, physiological cues such as blink rate and gaze shift inconsistencies are leveraged to enhance detection effectiveness. Experimental results on multiple datasets (Deepfake-and-Real Images, UADFV, FaceForensics++) demonstrate that our model achieves 97.61% accuracy, out performing standalone CNN’s. Grad-CAM visualization further enhances model interpretability, making the detection process more transparent. This research highlights the importance of combining CNN-based feature extraction with physiological-based detection for more robust and explainable deepfake forensics.