Attack definitions

Before explaining how we analyzed the technology, let us informally detail the different kinds of attacks that can be attempted against a watermarking scheme.

Let us call $A$ the original version of a song, and $AM$ the marked version. Obviously, recovering $A$ from $AM$, or something infinitely close to $A$ will remove the mark. However, it is not always necessary to perform such a hard task. It might sometimes be enough to produce a new song $C$ which is simply reasonably close to $A$, but which has the property that the mark cannot be detected in it anymore.

A simple example of this second type of attack would be to have musicians re-record the song. Clearly, the mark will not be present. The problem is to be able to produce something close enough to this original.

Actually, it is not even always necessary to ``remove'' the mark, it may be that the mark is still present, in some sense, but that the structure of the song has changed in such a way that the detector cannot find it anymore. This kind of attacks, that are sometimes called desynchronisation attacks, are known to be very efficient against images. While some papers have tried to explicitly fight them [PP99,JDJ99], geometric transformations on marked images make detection extremely hard if not impossible.

Let us now define the attacks that can be applied to defeat the proposed schemes. Our definitions are informal ones and should be seen from the practical point of view of an attacker.