Deep learning methods and advancements in digital image forensics

The volume of digital visual data is increasing dramatically year after year. At the same time, image editing has become easier and more precise. Malicious modifications are therefore more accessible. Image forensics provides solutions to ensure the authenticity of digital visual data.  Recognition of the source camera and detection of falsified images are among the main tasks. At first, the solutions were classical methods based on the artifacts produced during the creation of a digital image. Then, as in other areas of image processing, the methods moved to deep learning. First, we present a state-of-the-art survey of deep learning methods for image forensics. Our state-of-the-art survey highlights the need to apply pre-processing modules to extract artifacts hidden by image content. We also highlight the problems concerning image recognition evaluation protocols. Furthermore, we address counter-forensics and present compression based on artificial intelligence, which could be considered as an attack. In a second step, this thesis details three progressive evaluation protocols that address camera recognition problems. The final protocol, which is more reliable and reproducible, highlights the impossibility of state-of-the-art methods to recognize cameras in a challenging context. In a third step, we study the impact of compression based on artificial intelligence on two tasks analyzing compression artifacts: tamper detection and social network recognition. The performances obtained show on the one hand that this compression must be taken into account as an attack, but that it leads to a more important decrease than other manipulations for an equivalent image degradation. Future perspectives could be: i) the use of loss functions dedicated to pair similarity, such as contrastive of triplet; ii) the creation of a database including AI-based compression for the detection of double compression. The volume of digital visual data is increasing dramatically year after year. At the same time, image editing has become easier and more precise. Malicious modifications are therefore more accessible. Image forensics provides solutions to ensure the authenticity of digital visual data.  Recognition of the source camera and detection of falsified images are among the main tasks. At first, the solutions were classical methods based on the artifacts produced during the creation of a digital image. Then, as in other areas of image processing, the methods moved to deep learning. First, we present a state-of-the-art survey of deep learning methods for image forensics. Our state-of-the-art survey highlights the need to apply pre-processing modules to extract artifacts hidden by image content. We also highlight the problems concerning image recognition evaluation protocols. Furthermore, we address counter-forensics and present compression based on artificial intelligence, which could be considered as an attack. In a second step, this thesis details three progressive evaluation protocols that address camera recognition problems. The final protocol, which is more reliable and reproducible, highlights the impossibility of state-of-the-art methods to recognize cameras in a challenging context. In a third step, we study the impact of compression based on artificial intelligence on two tasks analyzing compression artifacts: tamper detection and social network recognition. The performances obtained show on the one hand that this compression must be taken into account as an attack, but that it leads to a more important decrease than other manipulations for an equivalent image degradation. Future perspectives could be: i) the use of loss functions dedicated to pair similarity, such as contrastive of triplet; ii) the creation of a database including AI-based compression for the detection of double compression.