<![CDATA[Deepfakes have become a serious threat to digital security, individual privacy, and the spread of disinformation globally. The main challenge in detecting manipulated media lies in balancing accuracy, model complexity, and generalization across varying levels of compression. This study proposes the E-SAFE (EfficientNet with Squeeze-and-Attention Feature Enhancement) model, a deepfake detection model integrating the EfficientNet-B0 architecture with the Squeeze-and-Excitation (SE) attention mechanism. This study adopted FaceForensics++ as a benchmark dataset for evaluating deepfake detection. The model was trained with the Adam optimizer and evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and Grad-CAM-based interpretability metrics. Experimental results indicated that E-SAFE attained 95% accuracy, 94% precision, 93% recall, 93% F1-score, and 98% ROC-AUC. The results surpassed the baseline EfficientNet-B0 while maintaining high computational efficiency. These results suggest that integrating the Squeeze-and-Excitation block enhanced the model's sensitivity to subtle facial manipulations without significantly increasing parameter complexity. The E-SAFE model has been shown to be superior in detecting subtle manipulations in deepfake images while maintaining parameter efficiency, thus potentially becoming a reliable solution for multimedia forensics.]]>
Copyrights © 2026
