Detecting Frame Deletion in Videos Using Supervised and Unsupervised Learning with Convolutional Neural Networks

In recent years, videos have been susceptible not only to any edition but also to a variety of forgeries. One of the most popular video forgeries is frame deletion, in which a group of frames is removed to hide specific actions from the human eye. When frame deletion occurs, videos selected as evidence lose their evidentiary value. This highlights the necessity of automation, especially for analyzing large volumes of videos. Thus, we measure the performance of two deep learning approaches for frame deletion detection. Both of them use Convolutional Neural Networks (CNN): The first one, a supervised 3DCNN model and, the second one, is an unsupervised model compound of VGG-16 and Resnet-50. We evaluated them using 10-fold cross-validation in the following datasets: UCF-101, VIFFD and DTD (Driving Test Dataset), which is our contribution to the data community. To the best of our knowledge, no comparison of both approaches using 10-fold cross-validation has been found in the literature before. Afterward, we analyze the results and make recommendations for future work in this area.