Ensemble deep learning fusion for detection of colorization based image forgeries

Abstract

Image forensics detects manipulation of digital images by tampering and counterfeiting process. While most works on Image forensics detect splicing, retouching and copy-move, very few have addressed colorization forgeries. Colorization or Fake colorization is a rapidly emerging area where colors of certain regions in image are manipulated with realistic colors. This is done maliciously to confound object recognition algorithms. Though some works are proposed to detect fake colorization, they can be deceived easily by introducing the pixel differences using statistical techniques. This work proposes a deep learning technique for detection of colorization forgeries which is resilient against deceiving attacks. Best set of discriminating features are extracted from Deep learning layers to recognize the differences in multiple channels of hue, saturation, value with aim to increase the accuracy of colorization forgery detection. Compared to most recent histogram based features, deep learning model is able to learn more intricate features about the distribution of intensity in hue, saturation, dark and value channels. Through experimental analysis, the proposed solution is found to provide at least 2% higher fake colorization detection accuracy compared to existing works