Cryptanalysis of Discrete-Sequence Spread Spectrum Watermarks

Assume that we are given a watermark (wm) embedding algorithm that performs well against generic benchmark-type attacks that comprise of simple operations that are independent of the algorithm and generally of the input as well. A natural question then is to ask for a nearly perfect cryptanalytic attack for the specific watermarking method. In this paper we present and analyze an attack on a state-ofthe-art Discrete-Sequence Spread Spectrum (dsss) audio watermarking algorithm. Our method uses detailed models for watermarked signal and almost always jams or recovers > 90% of the watermarking-key. It exploits the host and wm correlations, and the fact that one can locally correct errors in the wm estimates if the watermarking coefficients are discrete. It is natural to use error-correction codes in a watermarking algorithm, and we study the effects of the accompanying redundancy as well.