Research On Robust Digital Audio Watermarking

The need to protect audio works has become increasingly important due to the matureof audio compression techniques and the proliferation of the Internet. It is a verychallenging work to design a robust and secure watermarking algorithm in a hostileenvironment full of potential adversaries. This thesis is mainly concerning resolvingthe time-domain synchronization problems which are regarded the most challengingattacks to audio watermarking techniques. Also, it investigated the effect of errorcorrection coding on the baseline audio watermarking algorithm we proposed undervarious audio signal processing, synchronization attacks, and distortions introduced inStirmark for Audio. Random cropping (including jittering) and time-scale modification (TSM) are verytypical and challenging synchronization attacks to audio watermarking techniques.Compared with common audio signal processing attacks such as MP3 compression,low-pass filtering etc., they are more difficult to manage. To solve these openproblems, two series of novel algorithms were carefully designed and developed. Thefirst series of algorithms are aiming at finding some physical values on the discretewavelet transformation domain acting as "invariant watermark", which should berather stable under common audio signal processing and synchronization attacks.Based on the proposed algorithm, extensive experiments were performed toinvestigate the impact of error correction codes (Repetition code and BCH code) onimproving watermark detection efficiency. Several principles were given on how touse error correction codes more efficiently. The second series of algorithms are basedon localized watermarking technique. The watermark is repeatedly embedded intoselected embedding regions that represent music edges according to different methods.These regions usually correspond to percussion instruments like drums and can affectthe auditory quality of music greatly. Generally speaking, these regions will not bechanged much after undergoing audio signal processing and synchronization attacks. -iv-In this way, the watermark can not only be detected in regions that are almost thesame as the embedding ones but can also be expected to exempt from different kindsof distortions. By nature, this series of algorithms depend on the invariance propertyof these regions, i.e. implicit synchronization.