Signal Processing: Image Communication 89 (2020) 116008 Contents lists available at ScienceDirect Signal Processing: Image Communication journal homepage: www.elsevier.com/locate/image A robust JPEG compression detector for image forensics Chothmal Kumawat, Vinod Pankajakshan ∗ Department of Electronics & Communication Engineering, Indian Institute of Technology Roorkee, Roorkee 247667, India ARTICLE INFO Keywords: JPEG compression artifacts JPEG forensics JPEG anti-forensics Calibration ABSTRACT Identification of JPEG compressed images saved in uncompressed format (JPEG-U images) is an important issue in forensic analysis. The state-of-the-art JPEG compression detection methods fail to identify such images when subjected to post-processing/anti-forensic operations. In this paper, we propose a novel JPEG compression detector which is robust to post-processing and anti-forensic operations. The detector is based on the difference in the discrete cosine transform (DCT) coefficient distributions in the ac subbands of uncompressed images and JPEG-U images. We show theoretically and empirically that the probability of subband DCT coefficients which lie in the interval (−0.5, 0.5) is significantly different for a JPEG-U and the corresponding uncompressed image. This difference is exploited to derive a detection statistic which is compared with a threshold to detect JPEG-U images. The detector makes use of calibration, a technique used in steganalysis, to obtain the detection statistic. The experimental results show that the proposed detector significantly outperforms the state-of-the-art detectors, especially in the presence of post-processing and anti-forensic operations. 1. Introduction Due to the proliferation of low-cost digital cameras along with the recent advancements in Internet technologies, digital images are increasingly used for sharing visual information. Thousands of videos/ images are uploaded on social networking sites every second. With the widespread availability of easy-to-use image editing tools, it has now become very easy to modify the contents of digital images, even for non-professional users. Hence, it is important to verify the authenticity and integrity of digital images before trusting their content. Image forensic techniques can be used as an effective tool in this regard and have recently become an active area of research. Image forensic methods are broadly classified into two categories: passive and active forensic methods [1]. In active forensics, a known signature such as a digital watermark is inserted before disseminating the content. On the other hand, passive forensic methods use the intrinsic fingerprints left by the source camera or subsequent image processing operations. The non-intrusive nature of passive forensics makes it attractive in many practical scenarios. Digital images are stored either in an uncompressed or a com- pressed format. Among various compression standards, the JPEG is the most widely used lossy compression standard due to efficient data storage and low computational complexity. Hence, estimating the JPEG compression history is of special interest in forensic anal- ysis [2]. Various methods have been proposed for detecting JPEG compression [3–8], double JPEG compression [9–12] and for esti- mating JPEG quantization table [3,5,7,13]. These methods exploit ∗ Corresponding author. E-mail addresses: ckumawat@ec.iitr.ac.in (C. Kumawat), vinodfec@iitr.ac.in (V. Pankajakshan). various JPEG fingerprints like the histogram artifact, blocking artifact or truncation and rounding errors. Furthermore, various anti-forensic methods [14–18] have also been proposed to circumvent JPEG forensic detectors by suppressing the traces of compression artifacts. In response to anti-forensic methods, various anti-forensic detectors [19–25] are developed. These detectors exploit the artifacts introduced by the signal processing operations involved in the anti-forensic methods. A number of machine learning based image forensic techniques have been proposed in the recent years. Many of these techniques can also be used for detecting the presence of JPEG compression. These techniques can be categorized into two groups. One group consists of feature-based techniques in which a high-dimensional feature vector extracted from a test image is fed to a trained classifier. Many of these techniques [21, 22,24,26] borrow the features from the steganalysis literature. The second category is of deep learning based techniques [27–30]. As compared to other techniques, the machine learning based techniques can accurately classify images which had undergone a wide range of image processing operations. However, a drawback of such techniques is that they perform poorly in classifying the images that are subjected to any of the operations which were not considered during the training process. In this work, we address the problem of detecting the traces of pre- vious JPEG compression in images stored in an uncompressed format. Such a scenario arises in many practical situations. For instance, one may save a manipulated JPEG image in an uncompressed format to evade detection by JPEG forensic techniques. Another example is those https://doi.org/10.1016/j.image.2020.116008 Received 13 November 2019; Received in revised form 22 August 2020; Accepted 9 September 2020 Available online 18 September 2020 0923-5965/© 2020 Elsevier B.V. All rights reserved.