The avatars of noise in digital images and their use in image forensics

Images serve as potent information vectors, conveying a wealth of data and insights through visual representations. Their importance in various domains cannot be overstated, as they offer unique advantages for communication, understanding, and documentation. In an era characterized by the pervasive influence of digital imagery, image forensics represents a vital discipline that addresses the pressing need to uphold the veracity and trustworthiness of digital visual content. Images are naturally endowed with a fingerprint, embedded during the image formation process. Indeed, the creation of a digital image, spanning from its acquisition at the camera sensor to its final storage, imprints distinct artifacts that serve as a unique signature. The goal of this thesis is to retrieve this fingerprint through noise analysis. Along the camera processing pipeline, the initial Poisson noise is transformed by multiple operations tailored to each image formation process, leading to the final compressed image. As a consequence, noise residuals can yield significant forensic insights.Such cues allow forgery detection. Indeed, though nowadays manipulations have the capability to achieve a high degree of visual fidelity, they concurrently introduce alterations to the intrinsic structure of the image. Such disruptions in the inherent fingerprint are exploited by most forgery detection methods to spot tampered regions. The first part of this thesis focuses on this problem. Here, we propose two methods based on the detection of local inconsistencies of the noise model with respect to a background model. In particular, the Noisesniffer method adopts an a contrario validation step, aiming at controlling the expected number of false detections. We then explore the possibility of learning the forensic traces by means of deep convolutional networks instead of using hand-crafted features. Finally, this part ends with the evaluation of forgery detection methods themselves. We propose a methodology and a dataset to study the sensitivity of the detection tools to specific traces, as well as their ability to perform detection without semantic cues in the image.Source camera forensics tasks such source camera model identification or source device certification can also be achieved using the said fingerprint. Indeed, some of the forensic traces embedded during the image acquisition process are model-unique or device-unique. By isolating such signals, information about the source device can be obtained. The second part of this thesis focuses on these tasks. Here, we explore learning approaches to determine if a pair of images contain the same forensic traces. In addition, we propose a new statistical approach for source camera certification based on PRNU traces. Such an approach relies on two hypothesis tests based on local correlations which do not require computing empirical distributions.Still, nothing prevents the forgers from hiding the image fingerprint. This is why we devote the final part of this thesis to the analysis of different counter-forensics attacks. Highlighting the limitations of current forensic methods is important so that one can know how much trust can be put into an image and to encourage the exploration of alternative authentication methods. To this end, we analyze a novel approach recently introduced in the literature for camera trace erasing. This approach relies on an innovative hybrid loss for network training defined as a combination of three different losses: the embedded similarity loss, the truncated fidelity loss and the cross-identity loss. In addition, we propose a new counter-forensic attack based on diffusion models.