<p>The challenge of identifying the difference between genuine and altered images or videos is becoming increasingly more difficult due to advancements in deepfake technology. This research examines the application of two prominent deepfake detection models based on convolutional neural networks, EfficientNetB0 and ResNet50, using the Celeb-DF dataset. Both models utilize transfer learning, which employs pre-trained weights and requires training the model to detect minute details in the altered media. In addition to this, we propose an ensemble approach that combines predictions from multiple models to enhance detection performance through probability averaging. The results from the experiments show that EfficientNetB0 and ResNet50 have individual accuracies of 98.17% and 98.03% respectively, but the ensemble model achieved a more accurate 99.46%. This approach offers promising capabilities for immediate application in deepfake detection, particularly in the domains of forensic analysis of digital media, content safety verification systems, and cybersecurity. This research highlights the agility and effectiveness of modern model architectures in combating the rapidly evolving threat of synthetic media.</p>