Tardos Fingerprinting Codes in the Combined Digit Model

We formalize a new attack model for collusion secure codes, incorporating attacks on the underlying watermarking scheme as well as cut-and-paste attacks traditionally considered for collusion secure codes. We use this model to analyze the collusion resistance of two versions of the Tardos code, both for binary and nonbinary alphabets. The model allows us to consider different signal processing attacks on the content, namely the addition of noise and averaging attacks. The latter may result in content segments that have multiple watermarks embedded. We study two versions of the <formula formulatype="inline" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex Notation="TeX">$q$</tex></formula> -ary Tardos code in which the accusation method has been modified so as to allow for the detection of multiple symbols in the same content segment. We show that both variants yield efficient codes in the new model, parametrized for realistic attacker strengths.