A dual watermark-fingerprint system

Abstract

This article proposes a multimedia content protection system in which all copies of a protected object are identically watermarked, but each user has a distinct secret detection key that differs from the secret embedding key. An attacker with access to one detection key can fool the corresponding watermark detector but not other watermark detectors. Surprisingly, analogous to a criminal action, during this attack the attacker necessarily inserts his or her fingerprint into the modified content. Even a collusion clique of relatively large size cannot entirely remove the secret marks from the protected content by colluding their detection keys. More importantly, if the clique is not large enough, traces of the detection keys of all colluders can be detected with relatively high accuracy in the attacked clip. Our proposed watermark-fingerprint system achieves a minimum collusion size K that grows linearly with the size N of the marked object. In addition, we can augment our watermark-fingerprint system with a segmentation layer. The media content is partitioned into 5 segments, in which media players as well as forensic analyzers can reliably detect a watermark or fingerprint. Only detection keys that belong to the same segment can participate in the collusion clique. With segmentation, the minimum collusion size K grows as 0(N log N). Therefore, with or without segmentation, our watermark-fingerprint system significantly improves on the best-known asymptotic resistance to (fingerprint) collusion attacks of about O(N/sup 1/4/). Because we use a new protection protocol, comparing our system to classic fingerprint systems might seem unfair. However, such a comparison is important because the two technologies share a common goal: multimedia copyright enforcement. Our aim in this article is to characterize the collusion attacks against this system under the assumption that watermark detection is robust against signal-processing attacks on the protected object.