Deep Convolutional Neural Network for Object Forgery Detection in Video

Abstract

Talking of today’s digital revolution, where visual data is playing an imperative role, accessing, processing, and sharing of most of the information is typically attained with the help of video. These video sequences have shown their significance in various fields like news broadcasting, legal trials in court rooms, and many more but the doctoring of authentic visual content has made it uncertain to use as an evidence. Doctored video generation with a fast-growing rate done by easily accessible editing software like Adobe Photoshop, filmora, etc. have proved to be a major problem in maintaining its authenticity. The extent of forging is so vast that video spoofs reach our electronic-mail in-boxes, WhatsApp, Facebook or any other social media every minute and this fakery is totally indistinguishable that hence raise a demand for a new versatile field to perceive any alteration. Video forgery detection aims at restoring the trust and validating the authenticity by uncovering the counterfeits. But the traditional approaches used so far to detect forgeries have faced difficulties like less accurate detection rate and more false negatives. Nowadays, deep neural networks have been recognized as an effective technique in eradicating such troubles by learning significant features. The increasing attempt of video modification has drawn greater attention towards Deep Convolutional Neural Networks (DCNN) for achieving better counterfeits recognition.The proposed work is about “Deep Convolutional Neural Network for Object Forgery Detection in Video” that aims to detect forgery without requiring additional pre-embedded information of the frame. The proposed DCNN consists of various neurons where weights and biases are defined for individual neuron which helps the network to learn the data properly. Unlike other pre-existing learning-techniques, the proposed algorithm classifies the forged frames on the basis of correlation among them and the observed abnormalities using DCNN. The decoders used for batch normalization of input improves the training swiftness. It leads to an inordinate evidence in recognizing and discovering the fake regions. Simulation results are obtained on MATLAB 2018a with NVIDIA Cuda Graphics with REWIND and GRIP dataset which is rich in video inter-frame forgery effects. The outcomes so obtained with an average accuracy of 99% shows the superiority of the proposed algorithm as compared to existing one. The robustness of proposed algorithm is also tested on You Tube compressed video sequences. Recurrent Neural Networks can be combined with DCNN to achieve comparatively remarkable results in future.