Section: New Results

Image and video watermarking and security

Watermarking Robustness and Security

Participants : Caroline Fontaine, Fuchun Xie.

We have proposed several improvements of the Broken Arrows watermarking technique, strengthened its robustness and security. This work has been motivated by the BOWS-2 contest results; A. Westfeld proposed a quite efficient attack to cleanly remove the embedded watermark. This attack was based on a robustness weakness and a security attack strategy. The first step was to use linear regression to estimate the "original" (non watermarked) image from the watermarked one. We proposed a variant of Broken Arrows called AWC that counters this kind of robustness attack. The second step of Westfeld's attack was to observe a large set of watermarked images, to estimate corresponding original (non-watermarked) images through the afore-mentionned linear regression, then to estimate the embedded watermarks and use a clustering algorithm to classify them into bins. To attack a watermarked image, he substracted an average of the estimated watermarks lying into the corresponding bin. Recently, another security attack strategy has been proposed in the literature, using subspace estimation. We proposed a counter-measure to make both these attack fail.

Fingerprinting

Participants : Ana Charpentier, Caroline Fontaine.

Fingerprinting (also known as user forensics, traitor tracing, transactional watermarking, content serialization ...) aims at hiding in a robust and imperceptible way, an identifier of the consumer. The goal is to enable traceability of the content and to find back dishonest users who have illegally redistributed the content (for instance, posting it in a P2P network). Fingerprinting has a long history in research but real applications have slowly emerged this year: DRM systems over protect the content and thus they are not user-friendly. Content distribution could get rid off DRM thanks to fingerprinting used as a dissuasive weapon. This is a hot topic in the watermarking community. Fingerprinting is a difficult problem because it is a cross-design merging two layers: the fingerprinting code (the set of identifiers) and the watermarking technique hiding the identifiers in content.

A new trend in this domain is the probabilistic fingerprinting codes: The user identifiers are random binary strings with a secret statistic structure. G. Tardos introduced this concept in 2003, but it took time for the community to recognize his work as a major breakthrough. Tardos seminal work proposes a code design and proves that his performances are optimal. However, no clue is given on why it works so good, and on the setting of the parameters. Last year, we solved this point and proposed some improvements, deriving some optimal settings when the number of colluders is known, or when their collusion strategy is known. Following the steps of this previous work, in 2009, we have extended our results as follows:

• we proposed to integrate in the accusation process a dynamic estimation of the collusion strategy,

• and we provided more interesting accusation functions in the case where the collusion strategy.

Hence, we are able to provide a more powerful and reliable tracing technique. Our solution also enables us to reduce the length of the user identifiers which have to be embedded in the videos.