Imagerie forensique : un approche image-modèle

par Ludovic Darmet

Projet de thèse en Signal image parole telecoms

Sous la direction de François Cayre et de Kai Wang.

Thèses en préparation à Grenoble Alpes , dans le cadre de École doctorale électronique, électrotechnique, automatique, traitement du signal (Grenoble) , en partenariat avec Grenoble Images Parole Signal Automatique (laboratoire) et de Communication and Information in Compex Systems (CICS) (equipe de recherche) depuis le 01-11-2017 .


  • Résumé

    With the increasing popularity of digital cameras (especially those equipped on smartphones) and sophisticated photo editing software, it is now relatively easy, even for non-professionals, to create doctored images that do not reflect what happens in the reality [1]. Doctored images can have serious negative impacts on the society, e.g., misleading public opinion, deceiving consumers, and in particular hindering law enforcement if a doctored image is taken as a ‘‘convincing'' proof in court. Accordingly, many image forensic methods have been proposed to identify falsified images or to expose image's processing history. The basic idea of most forensic algorithms is to check, for a given image, the existence of inconsistency and/or deviation from intrinsic properties (e.g., geometrical, physical or statistical properties) of authentic images [2]. In the proposed thesis project, we aim to design reliable forensic detectors based on advanced image models. This appears to be a very promising line of research as our previous work [3] has shown that good forensic performance (i.e., in detecting routine image processing operations) can be achieved by using the relatively simple Gaussian Mixture Model (GMM) as generative model of small image patches of 8 by 8 pixels. More precisely, the main objectives of the proposed Ph.D. thesis are detailed below. First, for probabilistic image models [4] such as GMM, we will go beyond the simple likelihood comparison in [3] and try to derive a principled and theoretical framework for model parameter estimation which explicitly takes into account the image forensic performance. Second, for image generative models based on deep neural networks, e.g., the one proposed in the influential work of Goodfellow et al. [5], we will try to understand the underlying image model and utilize it for image forensic tasks, starting from the detection of routine image processing operations. Third, we want to go one step further, attempting to use both probabilistic and deep-network-based models for detecting visually plausible image forgeries, with a novel problem formulation and thorough experimental validation on large-scale datasets, e.g., the IEEE IFS challenge database [6]. At last, based on the insights gained from previous steps, we want to design new image models, either probabilistic or deep-network-based, which would be more adequate for image forensic tasks.

  • Titre traduit

    Digital image forensics: An image-model-based approach


  • Résumé

    With the increasing popularity of digital cameras (especially those equipped on smartphones) and sophisticated photo editing software, it is now relatively easy, even for non-professionals, to create doctored images that do not reflect what happens in the reality [1]. Doctored images can have serious negative impacts on the society, e.g., misleading public opinion, deceiving consumers, and in particular hindering law enforcement if a doctored image is taken as a ‘‘convincing'' proof in court. Accordingly, many image forensic methods have been proposed to identify falsified images or to expose image's processing history. The basic idea of most forensic algorithms is to check, for a given image, the existence of inconsistency and/or deviation from intrinsic properties (e.g., geometrical, physical or statistical properties) of authentic images [2]. In the proposed thesis project, we aim to design reliable forensic detectors based on advanced image models. This appears to be a very promising line of research as our previous work [3] has shown that good forensic performance (i.e., in detecting routine image processing operations) can be achieved by using the relatively simple Gaussian Mixture Model (GMM) as generative model of small image patches of 8 by 8 pixels. More precisely, the main objectives of the proposed Ph.D. thesis are detailed below. First, for probabilistic image models [4] such as GMM, we will go beyond the simple likelihood comparison in [3] and try to derive a principled and theoretical framework for model parameter estimation which explicitly takes into account the image forensic performance. Second, for image generative models based on deep neural networks, e.g., the one proposed in the influential work of Goodfellow et al. [5], we will try to understand the underlying image model and utilize it for image forensic tasks, starting from the detection of routine image processing operations. Third, we want to go one step further, attempting to use both probabilistic and deep-network-based models for detecting visually plausible image forgeries, with a novel problem formulation and thorough experimental validation on large-scale datasets, e.g., the IEEE IFS challenge database [6]. At last, based on the insights gained from previous steps, we want to design new image models, either probabilistic or deep-network-based, which would be more adequate for image forensic tasks.