Frame Deletion Detection in Videos Using Convolutional Neural Networks

With the broad adoption of digital media, videos are susceptible to various forms of forgery, making it crucial to ensure their authenticity, especially since they serve as digital evidence in contexts such as courts or forensic investigations. One of the main forgeries is frame deletion, which consists of removing frames from a video to hide specific actions from the human eye. Therefore, ways to automate and reduce errors when detecting frame deletion in videos are necessary, specially when analyzing a large volume of videos. We measure the performance of two Convolutional Neural Network (CNN) approaches for detecting frame deletion: a supervised 3DCNN model and an unsupervised model based on the VGG-16 architecture. We evaluated them in terms of accuracy, precision, recall and F1 score, using the following datasets: UCF-101, VIFFD and DTD (Driving Test Dataset), a dataset of authentic and forged driving test videos as our own contribution to the data community. Afterwards, we discuss the results and propose directions for future research in this area.