Abstract

The substantial increase in video recording and sharing in a short time, in addition to the easy-to-use editing tools, has created a strong necessity for efficient video forgery detection methods. The range of video manipulations, such as the usage of deep fakes, tampering of frames, and splicing, could endanger people's trust, cybersecurity, and the general stability of the digital media world. Previous works sometimes used temporal or spatial features separately, which was often not enough to accurately detect. This research paper proposes a new method based on 3D Convolutional Neural Networks (3D-CNNs) that captures the spatial as well as the temporal aspects of video frames simultaneously. The outlined architecture is trained on a face manipulation dataset, which includes real and fake videos, gathered from the Face Forensics repository. By distinguishing such minor areas of the video that have been changed, the system is capable of achieving high-performance levels in various forgery types. The first set consisted of two hundred videos, while the second set comprised three hundred videos. In the condition with the three hundred-video subsets, the Model demonstrated an accuracy level of 86%, precision of 0.87, recall of 0.86, and F1-score of 0.86 (weighted average). In the case of the two-hundred-video subsets, the test results were even better, with accuracy reaching 90%, precision being 0.92, recall 0.90, and F1-score 0.90. Compared with traditional classifiers like SVM, Random Forest, and LSTM models, the 3D-CNN method was always better judged by metrics such as accuracy, precision, and recall. These findings emphasize how powerful 3D-CNNs are in identifying spatiotemporal irregularities, thus enabling a more reliable and safer way to validate digital video content authenticity.