Deepfake generation techniques have advanced rapidly in recent years, posing significant risks for misinformation and privacy. In this paper, we propose an ensemble-based deepfake detection framework that leverages EfficientNet-B4 as the backbone architecture for both image and video analysis. By incorporating attention mechanisms and siamese training strategies, our system enhances feature discrimination and improves robustness against subtle manipulation artifacts. The framework is trained and evaluated on two widely adopted benchmarks—the DeepFake Detection Challenge (DFDC) and FaceForensics++ datasets. Experimental results demonstrate that the ensemble approach outperforms individual models, achieving higher accuracy and improved log-loss metrics, while also providing interpretability via attention maps. We further discuss the integration of temporal consistency analysis to better handle video data, and outline future directions for real-time deepfake forensic systems.