Non-Facial Video Spatiotemporal Forensic Analysis Using Deep Learning Techniques

Digital content manipulation software is working as a boon for people to edit recorded video or audio content. To prevent the unethical use of such readily available altering tools, digital multimedia forensics is becoming increasingly important. Hence, this study aims to identify whether the video and audio of the given digital content are fake or real. For temporal video forgery detection, the convolutional 3D layers are used to build a model which can identify temporal forgeries with an average accuracy of 85% on the validation dataset. Also, the identification of audio forgery, using a ResNet-34 pre-trained model and the transfer learning approach, has been achieved. The proposed model achieves an accuracy of 99% with 0.3% validation loss on the validation part of the logical access dataset, which is better than earlier models in the range of 90-95% accuracy on the validation set.