Détection de falsification d’images via l’analyse du démosaïquage : dévoilement d’une signature

Once considered reliable evidence, photographic images can no longer be assumed to depict the naked truth. With the advent of digital photography and the progress of photo editing tools, altering a picture has never been easier. While most of these modifications solely seek to enhance the image, they can potentially alter its very semantics. Concealing, modifying or adding a foreign object, all those can give an image a new and false meaning. Although these forgeries can easily be made visually realistic, they still distort the very fabric of the image. The formation of a digital image, from the camera sensors to storage, leaves traces, which act like a signature for the image. Modifying an image distorts these traces, creating detectable inconsistencies.Raw images are initially a mosaic of red, blue and green pixels. Missing colour values must be interpolated in a process known as demosaicing. In this thesis, we study the traces left by this process. The 2-periodic nature of the mosaic pattern leaves its imprint onto the image. Forgeries may dephase these traces, or even remove them entirely; mosaic pattern identification is consequently helpful in localizing tampered regions.Non-specific forgery detection methods can already analyse many traces in an image; nevertheless they remain blind to shifts in the mosaic, due to the translation-invariance of the convolutional neural networks on which most are based. Demosaicing-specific methods can thus provide complementary results for forgery detections. However, these have historically received little attention. Analysis of demosaicing artefacts is made harder by the vast array of often-undisclosed demosaicing algorithms, and above all by JPEG compression. Those artefacts, created early in the image formation pipeline and lying at the highest frequencies of the image, are quick to wane during compression.Yet, those artefacts can still be detected under mild compression. To channel the representative power of convolutional neural networks into the analysis of demosaicing artefacts, we introduce the notion of positional training. This self-supervised scheme trains the network to detect the modulo-2 position of each pixel, leveraging the translation invariance of convolution to make the network implicitly analyse demosaicing artefacts, its only clue to the modulo-2 position of a pixel. On top of that, internal training on a single potentially forged image can bolster the method’s robustness to JPEG compression on said image. Errors in the output of the neural network are then clues of mosaic inconsistencies. An a contrario paradigm then enables us to make automatic decisions on the authenticity of an image. Using only demosaicing artefacts, the proposed method beats the state of the art on several uncompressed datasets. On compressed images, it still provides decent results that are fully complementary with methods that are not mosaic-specific.Finally, we explore the very evaluation of forgery detection methods. We propose a methodology and dataset to study the sensitivity of forensic tools to specific traces, as well as their ability to make detections without semantic cues on the image. More than a simple evaluation tool, this methodology can be used to assess the strength and weaknesses of each method, as well as their complementarities.