DETECTION AND LOCALIZATION OF SPLICING FORGERY IN DIGITAL VIDEOS USING CONVOLUTIONAL AUTO-ENCODER AND GOTURN ALGORITHM

Abstract

In present days, individuals in the society are becoming increasingly relying on multimedia content, especially digital images and videos, to provide a reliable proof of the occurrence of events. However, the availability of powerful and user-friendly video editing tools makes it easy even for a novice to manipulate the content of a digital video which may be used as evidence during digital investigation. This has led to great concern regarding the authenticity of digital videos. Several techniques have been developed to detect video forgeries, but only few focused on video splicing forgery detection. However, those few techniques that focused on splicing forgery detection tend to depreciate in terms of accuracy of detection when the video is compressed. Therefore, it is important to devise new technique that can detect and localize splicing forgery for both compressed and uncompressed video. In this paper, a hybrid technique for the detection and localization of splicing forgery in both compressed and uncompressed digital videos using convolutional auto-encoder and Generic Object Tracking Using Regression Network (GOTURN) algorithm is proposed. The parameters of the auto-encoder are learned during the training phase on original video frames. During the testing phase, the auto-encoder reconstructs original frames with small reconstruction error and forged frames with large reconstruction error. The forged material is then tracked in subsequent frames using GOTURN algorithm. The result of the experiments demonstrates that the proposed detection technique can adequately detect video splicing with an Area Under the Receiver Operating Characteristics (AUROC) value of 0.9307