Xception-Based General Forensic Method on Small-Size Images

Abstract

Developing universal forensic methods that can simultaneously identify multiple image operations to identify the authenticity and processing history of an image has attracted more and more attention. Although numerous forensics tools and methods emerge to detect the traces left by various image operations, the accuracy with current techniques still decreased significantly as the size of the investigated images reduced. To overcome this issue, especially for small-size or highly compressed images, we propose a method using Xception-based convolution neural network. While CNNs-based methods are able to learn features directly from data for classification task, they are not well suited for forensic problems directly in their original form. Hence, we have added magnified layer in the preprocessing layer. The input images are magnified by the nearest neighbor interpolation algorithm in the magnified layer, which can preserve the property of image operations better than other magnified tools, and then input them into the CNN model for classification. Finally, we get adaptive average pooling function from global average pooling to adapt to any size of input pictures. We evaluate the proposed strategy on six typical image processing operations. Through a series of experiments, we show that this approach can significantly improve classification accuracy to 97.71% when the images are of size 64 × 64. More importantly, it outperforms all the existing general purpose manipulation forensic methods.