Security in watermarking and authentication of noisy data

Abstract

Many types of data, such as images and biometric templates, are noisy in the sense that their semantics and authenticity remain the same under small noises, which can be intentional. This leads to many research works on watermarking and authentication of noisy data, which focus mostly on the robustness against random noises. In this thesis, we study the security aspect of digital image watermarking and biometric authentication. Unlike many previous approaches where the attackers are random, we consider ``smart'' attackers, who can add carefully designed noises to the data to fail the system. We consider three scenarios in watermarking, with varying attacker capabilities: (1) Watermark detection with proxies, (2) Oracle attacks, and (3) Inversion attacks. We also consider secure fingerprint authentication, where the templates are modeled as noisy point-sets under a combination of different noises. We construct helper information schemes with provable security.