Discriminating multiple JPEG compressions using first digit features

Abstract

I . I N T R O D U C T I O NThe recent development of multimedia devices and editing tools, together with the proliferation of communication infrastructures and content sharing applications, has made the acquisition, the editing, and the diffusion of images and videos relatively-easy tasks. A single multimedia content downloaded from the internet can be the outcome of a long chain of processing steps. This fact introduces several concerns about the origin, the authenticity, and the trustability of images and videos downloaded from the network [1,2]. Moreover, identifying the origin and the authenticity (i.e., the absence of alterations after acquisition) of images proves to be a crucial element in court cases for the validation of evidences [3]. Moreover, detecting alterations permits inferring an objective quality evaluation of the analyzed multimedia content.From these premises, multimedia forensic analysts have been recently focusing on detecting alterations on images since they can be easily acquired and modified even by a non-expert user. In order to fulfill this task, many of the proposed works aim at identifying images, which have been compressed more than once estimating the coding parameters that characterize the coding stages that precede the last one [4]. This fact is justified by the observation that most Dipartimento di Elettronica, Informazione, e Bioingegneria (DEIB), Politecnico di Milano, P.za Leonardo da Vinci 32, 20133 Milano, ItalyCorresponding author: S. Milani Email: simone.milani@polimi.it of the digital multimedia contents are available in compressed format. Indeed, most of the images distributed over the internet are coded according to the JPEG standard [5].All the solutions proposed in the literature aim at detecting double compression on images assuming that some alterations can be performed between the first and the second compression stages. However, we believe that this assumption does not hold in many practical scenarios since analyzing the feasible processing chains for a given downloaded image it is deducible that more than two compression stages may have been applied. As an illustrative example, let us consider an image, which is originally compressed by the acquisition device (i.e., a video or photo camera) to be stored in the onboard memory. A second compression is performed by the owner, after editing the image to enhance the perceptual quality and adjust the format (e.g., brightness/contrast adjustment, rescaling, cropping, color correction, etc.). A third compression is performed whenever the content is uploaded to a blog or to an on-line photo album. As a matter of fact, it is reasonable to assume that a large number of digital images available online have gone through more than two compression stages performed by its owner, and could be further compressed by other users. In these cases, a method that identifies the number of compression stages proves to be extremely important in reconstructing the processing history of a content [6].This paper aims at identifying traces of...

DCT-domain deep convolutional neural networks for multiple JPEG compression classification

“…Detection of multiple JPEG compression has been addressed in [28,29]. Chu et al [30] established that using features from normalized DCT coefficient histogram of sub-band (2,3), maximum four number of compressions can be detected.…”

“…For generating the images compressed up to N times [29], a unique chain of compression is being generated as follows: the Quality factor of i th compression stage is randomly chosen from all possible values of QF i , which is described in Equation 2.…”

“…Value of dq min and dq max are chosen to be 6 and 12. These empirical numbers are adapted from [29].…”

“…The performance of the proposed method is compared with the method proposed in [29] with the value of QF N = 80. In [29], authors, have used 100 images from UCID dataset for training the SVM classifier and 200 images for testing purpose. Note that the confusion matrix reported here is directly taken from the author's paper.…”

“…Note that the confusion matrix reported here is directly taken from the author's paper. This Table 12 corresponds to Table Table 11: Confusion matrix for QF N = 85 C 1 C 2 C 3 C 4 C 5 C 1 99.20 0.40 0.40 0.00 0.00 C 2 0.44 98.82 0.00 0.74 0.00 C 3 1.34 0.14 98.50 0.02 0.00 C 4 0.00 0.10 0.00 99.90 0.00 C 5 0.00 0.02 0.10 0.02 99.86 2(a) of the author's paper in [29]. For the comparison purpose, we have also trained the model with 100 images and tested with 200 images.…”