Abstract
These days, videos can be easily recorded, altered and shared on social and electronic media for deception and false propaganda. However, due to sophisticated nature of the content alteration tools, alterations remain inconspicuous to the naked eye and it is a challenging task to differentiate between authentic and tampered videos. During the process of video tampering the traces of objects, which are removed or modified, remain in the frames of a video. Based on this observation, in this study, a new method is introduced for discriminating authentic and tampered video clips. This method is based on deep model, which consists of three types of layers: motion residual (MR), convolutional neural network (CNN), and parasitic layers. The MR layer highlights the tampering traces by aggregation of frames. The CNN layers encode these tampering traces and are learned using transfer learning. Finally, parasitic layers classify the video clip (VC) as authentic or tampered. The parasitic layers are learned using an efficient learning method based on extreme learning theory; they enhance the performance in terms of efficiency and accuracy. Intensive experiments were performed on various benchmark datasets to validate the performance and the robustness of the method; it achieved 98.89% accuracy. Comparative analysis shows that the proposed method outperforms the state-of-the-art methods.
Highlights
In the digital era of 21st century, mobile phones, personal digital assistants (PDAs) and digital camcorders are available to acquire the videos
The proposed method based on the deep model gives better accuracy (98.89%) and efficiency for the classification of authentic and tampered video clip (VC) on different datasets than the state-of-the-art methods
The objective of this study is to investigate the capability of deep convolutional neural network (CNN) for classification of authentic and tampered videos and target to answer these questions: (a) Which model of deep learning should be employed? (b) Which part of the model should be used for feature extraction and representation of the tampered traces? (c) How many layers should be transferred in order to obtain best performance? (d) Which classifier is best to classify authentic and tampered videos? (e) How to modify the existing deep CNN model to make a new model for classification of authentic and tampered videos?
Summary
In the digital era of 21st century, mobile phones, personal digital assistants (PDAs) and digital camcorders are available to acquire the videos. These videos can be redistributed for different purposes like video conferences, surveillance systems, information propagation to the media houses and social media websites. The quality and contents of the videos can be modified with different video editing tools. Different objects can be added, removed or replaced within a video frame or a series of frames, whereas in the temporal domain, a number of frames are added, removed or replaced from the video [4].
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have