Multimedia Protection using Content and Embedded Fingerprints

Abstract

Improved digital connectivity has made the Internet an important medium for multimedia distribution and consumption in recent years. At the same time, this increased proliferation of multimedia has raised significant challenges in secure multimedia distribution and intellectual property protection. This dissertation examines two complementary aspects of the multimedia protection problem that utilize content fingerprints and embedded collusion-resistant fingerprints. The first aspect considered is the automated identification of multimedia using content fingerprints, which is emerging as an important tool for detecting copyright violations on user generated content websites. A content fingerprint is a compact identifier that captures robust and distinctive properties of multimedia content, which can be used for uniquely identifying the multimedia object. In this dissertation, we describe a modular framework for theoretical modeling and analysis of content fingerprinting techniques. Based on this framework, we analyze the impact of distortions in the features on the corresponding fingerprints and also consider the problem of designing a suitable quantizer for encoding the features in order to improve the identification accuracy. The interaction between the fingerprint designer and a malicious adversary seeking to evade detection is studied under a game-theoretic framework and optimal strategies for both parties are derived. We then focus on analyzing and understanding the matching process at the fingerprint level. Models for fingerprints with different types of correlations are developed and the identification accuracy under each model is examined. Through this analysis we obtain useful guidelines for designing practical systems and also uncover connections to other areas of research. A complementary problem considered in this dissertation concerns tracing the users responsible for unauthorized redistribution of multimedia. Collusion-resistant fingerprints, which are signals that uniquely identify the recipient, are proactively embedded in the multimedia before redistribution and can be used for identifying the malicious users. We study the problem of designing collusion resistant fingerprints for embedding in compressed multimedia. Our study indicates that directly adapting traditional fingerprinting techniques to this new setting of compressed multimedia results in low collusion resistance. To withstand attacks, we propose an anti-collusion dithering technique for embedding fingerprints that significantly improves the collusion resistance compared to traditional fingerprints.