Publications
For a complete list of publications, go to Google Scholar.
A Reliable JPEG Quantization Table Estimator
IPOL, May 2022
JPEG compression is a commonly used method of lossy compression for digital images. The degree of compression can be adjusted by the choice of a quality factor QF. Each software associates this value to a quantization table, which is an 8 x 8 matrix used to quantize the DCT coefficients of an image. We propose a method for recovering the JPEG quantization table relying only on the image information, without any metadata from the file header; thus the proposed method can be applied to an uncompressed image format to detect a previous JPEG compression. A statistical validation is used to decide whether significant quantization traces are found or not, and to provide a quantitative measure of the confidence on the detection.
Multimedia Security 1: Authentication and Data Hiding: Chapter 2: How to Reconstruct the History of a Digital Image, and of Its Alterations
John Wiley & Sons, March 2022
This chapter reviews the operations undergone by the raw image and describes the artifacts they leave in the final image. The Internet, digital media, new means of communication and social networks have accelerated the emergence of a connected world where perfect mastery over information becomes utopian. Image manipulation can serve the interests of criminal or terrorist organizations as part of their propaganda. Noise estimation is a necessary preliminary step to most image processing and computer vision algorithms. Noise inconsistency analysis is a rich source for forgery detection due to the fact that forged regions are likely to present different noise models from the rest of the image. Image demosaicing leaves artifacts that can be used to find falsifications. The chapter seeks to determine the compression history of an image. It focuses on the JPEG algorithm, which is nowadays the most common method to store images.
Non-Semantic Evaluation of Image Forensics Tools: Methodology and Database
IEEE WACV, January 2022
We propose a new method to evaluate image forensics tools, that characterizes what image cues are being used by each detector. Our method enables effortless creation of an arbitrarily large dataset of carefully tampered images in which controlled detection cues are present. Starting with raw images, we alter aspects of the image formation pipeline inside a mask, while leaving the rest of the image intact. This does not change the image's interpretation; we thus call such alterations" non-semantic", as they yield no semantic inconsistencies. This method avoids the painful and often biased creation of convincing semantics. All aspects of image formation (noise, CFA, compression pattern and quality, etc.) can vary independently in both the authentic and tampered parts of the image. Alteration of a specific cue enables precise evaluation of the many forgery detectors that rely on this cue, and of the sensitivity of more generic forensic tools to each specific trace of forgery, and can be used to guide the combination of different methods. Based on this methodology, we create a database and conduct an evaluation of the main state-of-the-art image forensics tools, where we characterize the performance of each method with respect to each detection cue.
ZERO: a Local JPEG Grid Origin Detector Based on the Number of DCT Zeros and its Applications in Image Forensics
IPOL, December 2021
This work describes a method for detecting JPEG compression as well as its grid origin. The JPEG algorithm performs a quantization of the DCT coefficients of non-overlapping 8×8 blocks of images, setting many of those coefficients to zero. The method described here exploits these facts and identifies the presence of a JPEG grid when a significant number of DCT zeros is observed for a given grid origin. This method can be applied globally to identify a JPEG compression, and also locally to identify image forgeries when misaligned or missing JPEG grids are found. The algorithm includes a statistical validation step according to Desolneux, Moisan and Morel’s a contrario theory, which associates a number of false alarms (NFA) with each tampering detection. Detections are obtained by a threshold of the NFA, which renders the method fully automatic and endows it with a false alarm control mechanism.
Sécurité Multimédia 1, Chapitre 1 : Comment reconstruire l’histoire d’une image digitale, et de ses altérations ?
ISTE, June 2021
Plus de 80 % des données transmises sur les réseaux et archivées dans nos ordinateurs, tablettes, téléphones portables ou sur les Clouds sont des données multimédia (images, vidéos, son, données 3D) pour des applications allant du jeu vidéo aux données médicales, en passant par la conception assistée par ordinateur, la vidéosurveillance et la biométrie. Il devient urgent de sécuriser ces données multimédia, que ce soit pendant leur transmission, leur archivage, mais également lors de leur visualisation. En effet, avec le « tout numérique », il devient de plus en plus facile de les copier, de les visualiser sans droit, de se les approprier ou de les falsifier. Sécurité multimédia 1 analyse les questions d’authentification de données multimédia, des codes et de l’insertion de données cachées, du côté défenseur comme du côté attaquant. Concernant l’insertion de données cachées, il traite également les aspects invisibilité, couleur, traçage et données 3D, tout comme la détection de message caché dans une image par stéganalyse.
Local JPEG Grid Detector via Blocking Artifacts, a Forgery Detection Tool
IPOL, April 2020
Image JPEG compression leaves blocking artifact traces. This paper describes an algorithm that exploits those traces to locally recover the grid embedded in the image by the JPEG compression. The algorithm returns a list of grids associated with different parts of the image. The method uses Chen and Hsu’s cross-difference to reveal the artifacts. Then, an a contrario validation step according to Desolneux, Moisan and Morel’s theory delivers for each detected grid a Number of False Alarms (NFA) which tells how unlikely it is that the detection is due to chance. The only parameter is the step size of the windows used, which represents the exhaustiveness of the method. The application to image forgery detection is twofold: first, the presence of discrepant JPEG grids with low NFA is a strong forgery cue; second, knowledge of the grid is anyway required for further JPEG forensic analysis.
Détection de grille JPEG par compression simulée
GRETSI, August 2019
La rétro-ingénierie est un outil efficace pour détecter les falsifications d’images. Nous proposons une méthode simple et fiable pour détecter la compression JPEG et la position de sa grille, et la comparons à l’état de l’art. La méthode peut être utilisée afin de déterminer si une image a été rognée, ce qui est souvent le premier indice de falsification.
JPEG Grid Detection based on the Number of DCT Zeros and its Application to Automatic and Localized Forgery Detection
CVPR Workshop, June 2019
This work proposes a novel method for detecting JPEG compression, as well as its grid origin, based on counting the number of zeros in the DCT of 8 × 8 blocks. When applied locally, the same method can be used to detect grid alignment abnormalities. It therefore detects local image forgeries such as copy-move. The algorithm includes a statistical validation step which gives theoretical guarantees on the number of false alarms and provides secure guarantees for tampering detection. The performance of the proposed method is illustrated with both quantitative and visual results from well-known image databases and comparisons with state of the art methods.
Parallel-beam ROI reconstruction with differentiated backprojection and angularly subsampled complementary sinograms
Fifth International Conference on Image Formation in X-Ray Computed Tomography, May 2018
Recently, we introduced a parallel-beam two-pass analytical reconstruction that allows truncation to be accounted for in the image domain rather than the projection domain. In particular, we showed that backprojection of a vastly angularly undersampled sinogram of un-truncated data could be used to extrapolate the backprojection of a finely sampled, fully truncated sinogram of the same object to perform more accurate region-of-interest (ROI) imaging. The same extrapolation idea can be performed using differentiated backprojection (DBP). The goal of this study is to give a general DBP-based formula when reconstructing a finite set of projections in parallel geometry. We discuss the discretization of this formula, in particular when the image grid size is large with respect to the number of projections, and we show how it can be applied to our extrapolation problem.