Liveness detection in facial recognition systems is essential to counteract spoofing attacks, such as those involving photos, videos, or masks. Traditional methods, including hardware-based solutions like infrared cameras and software based approaches analyzing visual cues, often face challenges related to cost, complexity, and susceptibility to advanced spoofing techniques. Recent advancements in deep learning, particularly in the use of Recurrent Neural Networks (RNNs) such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks for capturing temporal dependencies, have shown promise in enhancing liveness detection. This study compares the performance of LSTM and GRU models for video-based liveness detection, highlighting their respective strengths and limitations. The LSTM model achieved a final validation accuracy of 87.5%, with a balanced precision, recall, and F1-score, and an overall accuracy of 88%. The GRU model achieved a final validation accuracy of 84.38%, with a slightly higher precision, recall, and F1-score for the positive class, and an overall accuracy of 84%. These results demonstrate that while LSTM models provide higher accuracy and better generalization, GRU models offer computational efficiency and faster training times. The findings underscore the potential of deep learning models to significantly improve the robustness and reliability of facial recognition systems against spoofing attacks.