La vie secrète des images JPEG : Détection de falsification via les traces de compression

With the advent of digital photography and the progress of photo editing tools, modifying an image has become easy and accessible to all. Most of these modifications aim at improving the image, but some are intended at altering its meaning. Such forgeries can easily be made visually realistic. Fortunately, they also distort the very fabric of the image, as it is formed during the camera pipeline. Indeed, the formation of a digital image, from camera sensors to storage in JPEG format, leaves characteristic artifacts, which act as a signature of the image. Concealing, modifying or adding a foreign object in the image distorts these signatures and creates detectable inconsistencies.This thesis studies JPEG compression and the fingerprint (in the form of 8x8 block patterns) that it leaves on the image. The patterns are then exploited to propose several image forensic algorithms. These methods are based on the a contrario statistical theory, leading to automatic decision algorithms, not requiring the need for visual interpretation. Among the proposed methods is ZERO, a new image forensic algorithm which analyzes JPEG artifacts and detects image tampering when a local anomaly is found to be statistically significant. This method is the current state of the art in forgery detection by JPEG traces and can be used by the general public through a tool developed by Agence France Presse.The thesis also explores the evaluation of forgery detection methods in general. A new non-semantic methodology is proposed, together with the associated datasets. They enable to characterize the sensitivity of forensic tools relative to specific image traces and to avoid the difficulties of semantic evaluations.