The rise in deepfake technology is encouraging the creation of deepfake images. It is getting difficult for traditional methods to detect manipulated images with a high level of accuracy as they rely on static analysis and lack effective means of capturing temporal inconsistencies. In this research, we have compared four architectures—Convolutional Neural Networks, Recurrent Neural Networks, Hybrid CNN-LSTM, and CNN-RNN on a dataset containing real and fake images. Among these four models, the hybrid CNN-LSTM architecture achieved the highest accuracy, effectively capturing both spatial and temporal dependencies for image classification. The primary objective of the research is to improve detection accuracy by capturing sequential anomalies in deepfake images. A comprehensive data augmentation strategy was implemented to enhance model robustness. The model was trained and tested on a dataset from Kaggle consisting of 5000 real and 5000 fake images. We have applied preprocessing techniques such as resizing and normalization to standardize input data. The model achieved a validation accuracy of 95% using a hybrid CNN-LSTM model.