An Estimation Attack on Content-Based Video Fingerprinting

In this paper we propose a simple signal processing procedure that aims at removing low-frequency fingerprints embedded in video signals. Although we construct an instance of the attack and show its efficacy using a specific video fingerprinting algorithm, the generic form of the attack can be applied to an arbitrary video marking scheme. The proposed attack uses two estimates: one of the embedded fingerprint and another of the original content, to create the attack vector. This vector is amplified and subtracted from the fingerprinted video sequence to create the attacked copy. The amplification factor is maximized under the constraint of achieving a desired level of visual fidelity. In the conducted experiments, the attack procedure on the average halved the expected detector correlation compared to additive white gaussian noise. It also substantially increased the probability of a false positive under attack for the addressed fingerprinting algorithm.