<p>Deepfake technology is evolving, and its consequences are extensive, posing major risks to a variety of industries, including video surveillance systems. The growing amount of deepfake videos undermines the reliability of CCTV footage, leading to compromised evidence in criminal investigations. Existing deepfake detection methods are CNN-based and depend on benchmark datasets that do not represent real-world contexts, resulting in poor robustness when applied outside of controlled environments. This exposes a critical gap in trustworthiness of existing detection models, especially in high-stakes domains such as forensic analysis. In order to address these limitations, this study proposes a deepfake detection method that employs ConvLSTM and LRCN models to account for the temporal context inherent in video data. To facilitate this research, a new dataset, CCTV-DF, is created, consisting of 1000 real and fake CCTV videos, taken from realistic surveillance environments. To analyze model robustness, the dataset was exposed to a variety of video degradation procedures, including noise addition, compression, and grayscale conversion. These perturbations were designed to simulate typical issues encountered in real-world CCTV footage, such as poor quality and distortions caused by surroundings. The results show that when trained on novel CCTV-DF dataset, the ConvLSTM and LRCN models performed very well in detecting deepfakes in real-world circumstances. Specifically, ConvLSTM model achieved 99.8% accuracy and LRCN model achieved an accuracy of 99%. This study contributes to the field of deepfake detection by providing more reliable and robust approach for detecting deepfake videos within the context of video surveillance systems.</p>