Abstract
This paper presents the development, training, and performance evaluation of a customized 4-layer Convolutional Neural Network (CNN) for classifying deepfake and real images. The model was trained on Google Colab, using a publicly available dataset. Data augmentation techniques were applied to increase the size and diversity of the training dataset. The resulting accuracy on the test dataset was 97.5 %, indicating strong performance in distinguishing manipulated images from real ones. A comparison with other existing deepfake detection models, including DenseNet, Xception, EfficientNet, and a similar 5-layer CNN, revealed the superior accuracy and runtime efficiency (4m 44s/epoch) of the proposed CNN. Additionally, the proposed model was adapted for video classification, by implementing a custom designed algorithm based on analyzing each 10th frame of a video and averaging the predictions to determine a final verdict on the video’s authenticity. The results suggest that the customized 4-layer CNN outperforms predefined models, offering a promising solution for real-time deepfake detection tasks.