Prediction of Multiclass Tampering Using Advanced Hybrid Neural Networks

Abstract

Image tampering is now common practice with the use of high-performance hardware and state of art image modifying software. Social media provide various ways to share the fake news without any confirmation from the source. Fake news may exist in multiple forms like text, images and videos. The reliability and reputation of any person or organization can easily be affected with fake news. There are different ways to tamper the images which are used in common practices. Recently, Researchers are found to be very keen of using the deep learning techniques for detecting the fake images. This paper presents the approach by developing model usingadvancedneural networks to detect the type of tampering done on the images.We have put our efforts on the hybrid deep learning models to train on tampered images by categorizing the image datasets.TDC (Tampered Data Classes) module has been developed using multiple python scripts which segregates the images into 4 classesas Original, Splicing, Region Removal and Retouching images. Neural networks such as XceptionNet, InceptionNet, and RestNet have been used in hybrid mode as pre-trained CNN models to perform the feature extraction and classification. Highest accuracy has been achieved by the hybrid models of XceptionNet, InceptionNet and ResNet with our image datasets.To predict the multiclass classification about the tampered type the convolutional and pooling layers have been altered to obtain multiple features for desired classification. Several experiments have been done to test the proposed method. The experiments outcome reveals that the proposed method provides significant improvements, in identifying the tampered class such as Splicing, Region Removal, andretouching forgery.The performance of each model in terms of Loss, Accuracy, Validation Loss and Validation Accuracy has been generated which indicates that Hybrid Net has achieved the highest validation accuracy with our dataset. TheValidation accuracy is 94 %and the training accuracy is (96 %) with validation loss of 15 % for the Hybrid Model. This framework provides high accuracy as Block chain verifies all the transaction on the images and this verified data is supplied into the various layers of advanced CNNs.