Institute of Information Science, Academia Sinica

The need to compress signals (audio, video, etc.) into short bitstrings for the purpose of reliable identification (instead of reconstruction) arises in many modern applications, including management of user-submitted content websites such as YouTube; connected audio, e.g. the Shazam application running on smartphones; and contextual advertising. These short bitstrings are called robust hashes, or fingerprints of the signal. A successful application of content identification requires that different versions of a signal can be matched reliably based on their fingerprints (a property called robustness) whereas unrelated signals yield a negative match (this property is called discriminability). This talk will overview the fundamental concepts behind these applications and outline practical approaches.