Tatouage d'images : gain en robustesse et intégrité des images

Digital watermarking consists in hiding subliminal information (i.e. invisible or inaudible) in a document to ensure security services (e.g. copyright protection, image authentication, non repudiation, etc.). Compare to other techniques (e.g. adding simply the data in the header of the file), the message is robustly embedded in the data itself. Therefore the watermark is theoretically independent of the file format and it can be detected or extracted even if the image has undergone some manipulations. In this dissertation we describe a robust watermarking algorithm based on self-similarities derived from fractal coding. We mainly focus on the formatting step of the watermark, i.e. on the way the binary information is arranged before being inserted in the image. This operation aims at adapting the data to the channel (i.e. the image) and to encrypt the information to prevent its extraction by an unauthorized person. Among the formatting operations, we detail those which consist in adding redundancy (i.e. error correcting codes) and additional bits to overcome malicious or not image manipulations. We describe two resynchronization techniques that compensate for geometric deformations. The first method is dedicated to limited local or global geometric distortions (c.f. Stirmark). The second technique is tailored to compensate for global linear transformations such as rotation, scaling and shearing. Finally, we concentrate on watermarking-based image authentication. The first method is based on a robust watermarking scheme. The basic idea is to hide into the image features on itself. During the verification step, this information is extracted and compared to the characteristics of the tested image. The differences indicate which regions of the image have been altered. The main advantage of this method relies in the self-sufficiency of the image: no additional information is required. In a second approach the characteristics are not embedded in the image but stored in a third-party database. This information can only be retrieved with an ID embedded into the image. Although the use of this approach is restricted by the third-party requirement, it offers a more reliable service.