Multiclass Classification of Authentic and Forged Images Based on Deep Learning

Digital manipulation of images is commonly done by the use of advanced tools and technologies. As these manipulations cannot be easily seen by naked eyes, it is very difficult to differentiate the original and forged images. Image splicing and copy-move manipulations are the most common techniques to generate such manipulated images called as forged images. Image forgery can be detected and localized based on most important features called as descriptor of an image. The manipulations can be identified by correlation among nearby pixels using various deep learning-based models. A Convolutional Neural Network is applied to identify the related discrepancies. The convolution operation produced the output feature map, which is a meaningful information related to discrepancies. To detect copy-move and image splicing forgeries, a deep learning-based method is proposed here. The images are classified into three classes. Forged images can be Copy-move or Spliced and third class is for Authentic images. CASIA_v2 dataset is used for training and testing purposes. The various performance parameters like accuracy, F1-score, precision, recall are computed and compared for three types of images. The obtained accuracy is 93.54% for this multi-class classification with maximum precision and F1-score of 95.69 and 96.30%, respectively.