Digital audio watermarking robust against Locality Sensitive Hashing

Audio fingerprinting is an audio feature vector and a technique to identify the music source by comparing with original fingerprints in a fingerprint database. The fingerprints are often hashed to shorten codes. The content similarities between fingerprints are maintained in the hashed code. Such a hash function family is called Locality Sensitive Hashing (LSH). Because audio fingerprint using LSH is known to be resistant against CPO (Content Preserving Operations, perceptively acceptable manipulations) such as compression, noise adding, mean filtering, it is possible to identify the original source even if the source was slightly modified. On the other hand, mixed arrangement (mashup) of several music sources is allowed as legitimate artistic expression. In the conventional fingerprint based retrieval system, the mixed arrangements could identify the origins segmentally but the arranger's authorization is ignored. In this report, we propose an audio watermarking method robust against LSH coding. That is, the arranger information is watermarked in the audio signal and it is detectable from not only stego audio signal but also stego audio fingerprint of LSH.